Related
I have a C++ program running on Linux in which a new thread is created to do some computationally expensive work independent of the main thread (The computational work completes by writing the results to files, which end up being very large). However, I'm getting relatively poor performance.
If I implement the program straightforward (without introducing other threads), it completes the task in roughly 2 hours. With the multi-threaded program it takes around 12 hours to do the same task (this was tested with only one thread spawned).
I've tried a couple of things, including pthread_setaffinity_np to set the thread to a single CPU (out of the 24 available on the server I'm using), as well as pthread_setschedparam to set the scheduling policy (I've only tried SCHED_BATCH). But the effects of these have so far been negligible.
Are there any general causes for this kind of problem?
EDIT: I've added some example code that I'm using, which is hopefully the most relevant parts. The function process_job() is what actually does the computational work, but it would be too much to include here. Basically, it reads in two files of data, and uses these to perform queries on an in-memory graph database, in which the results are written to two large files over a period of hours.
EDIT part 2: Just to clarify, the problem is not that I want to use threads to increase the performance of an algorithm I have. But rather, I want to run many instances of my algorithm simultaneously. Therefore, I expect the algorithm would run at a similar speed when put in a thread as it would if I didn't use multi-threads at all.
EDIT part 3: Thanks for the suggestions all. I'm currently doing some unit tests (seeing which parts are slowing down) as some have suggested. As the program takes a while to load and execute, it is taking time to see any results from the tests and therefore I apologize for late responses. I think the main point I wanted to clarify is possible reasons why threading could cause a program to run slowly. From what I gather from the comments, it simply shouldn't be. I'll post when I can find a reasonable resolution, thanks again.
(FINAL) EDIT part 4: It turns out that the problem was not related to threading after all. Describing it would be too cumbersome at this point (including the use of compiler optimization levels), but the ideas posted here were very useful and appreciated.
struct sched_param sched_param = {
sched_get_priority_min(SCHED_BATCH)
};
int set_thread_to_core(const long tid, const int &core_id) {
cpu_set_t mask;
CPU_ZERO(&mask);
CPU_SET(core_id, &mask);
return pthread_setaffinity_np(tid, sizeof(mask), &mask);
}
void *worker_thread(void *arg) {
job_data *temp = (job_data *)arg; // get the information for the task passed in
...
long tid = pthread_self();
int set_thread = set_thread_to_core(tid, slot_id); // assume slot_id is 1 (it is in the test case I run)
sched_get_priority_min(SCHED_BATCH);
pthread_setschedparam(tid, SCHED_BATCH, &sched_param);
int success = process_job(...); // this is where all the work actually happens
pthread_exit(NULL);
}
int main(int argc, char* argv[]) {
...
pthread_t temp;
pthread_create(&temp, NULL, worker_thread, (void *) &jobs[i]); // jobs is a vector of a class type containing information for the task
...
return 0;
}
If you have plenty of CPU cores, and have plenty of work to do, it should not take longer to run in multithreaded than single threaded mode - the actual CPU time may be a fraction longer, but the "wall-clock time" should be shorter. I'm pretty sure that your code has some sort of bottleneck where one thread is blocking the other.
This is because of one or more of these things - I'll list them first, then go into detail below:
Some lock in a thread is blocking the second thread from running.
Sharing of data between threads (either true or "false" sharing)
Cache thrashing.
Competition for some external resource causing thrashing and/or blocking.
Badly designed code in general...
Some lock in a thread is blocking the second thread from running.
If there is a thread that takes a lock, and another thread wants to use the resource that is locked by this thread, it will have to wait. This obviously means the thread isn't doing anything useful. Locks should be kept to a minimum by only taking the lock for a short period. Using some code to identify if locks are holding your code, such as:
while (!tryLock(some_some_lock))
{
tried_locking_failed[lock_id][thread_id]++;
}
total_locks[some_lock]++;
Printing some stats of the locks would help to identify where the locking is contentious - or you can try the old trick of "Press break in the debugger and see where you are" - if a thread is constantly waiting for some lock, then that's what's preventing progress...
Sharing of data between threads (either true or "false" sharing)
If two threads use [and update the value of it frequently] the same variable, then the two threads will have to swap "I've updated this" messages, and the CPU's have to fetch the data from the other CPU before it can continue with it's use of the variable. Since "data" is shared on a "per cache-line" level, and a cache-line is typically 32-bytes, something like:
int var[NUM_THREADS];
...
var[thread_id]++;
would classify as something called "false sharing" - the ACTUAL data updated is unique per CPU, but since the data is within the same 32-byte region, the cores will still have updated the same are of memory.
Cache thrashing.
If two threads do a lot of memory reading and writing, the cache of the CPU may be constantly throwing away good data to fill it with data for the other thread. There are some techniques available to ensure that two threads don't run in "lockstep" on which part of cache the CPU uses. If the data is 2^n (power of two) and fairly large (a multiple of the cache-size), it's a good idea to "add an offset" for each thread - for example 1KB or 2KB. That way, when the second thread reads the same distance into the data region, it will not overwrite exactly the same area of cache that the first thread is currently using.
Competition for some external resource causing thrashing and/or blocking.
If two threads are reading or writing from/to the hard-disk, network card, or some other shared resource, this can lead to one thread blocking another thread, which in turn means lower performance. It is also possible that the code detects different threads and does some extra flushing to ensure that data is written in the correct order or similar, before starting work with the other thread.
It is also possible that there are locks internally in the code that deals with the resource (user-mode library or kernel mode drivers) that block when more than one thread is using the same resource.
Generally bad design
This is a "catchall" for "lots of other things that can be wrong". If the result from one calculation in one thread is needed to progress the other, obviously, not a lot of work can be done in that thread.
Too small a work-unit, so all the time is spent starting and stopping the thread, and not enough work is being done. Say for example that you dole out small numbers to be "calculate if this is a prime" to each thread, one number at a time, it will probably take a lot longer to give the number to the thread than the calculation of "is this actually a prime-number" - the solution is to give a set of numbers (perhaps 10, 20, 32, 64 or such) to each thread, and then report back the result for the whole lot in one go.
There are plenty of other "bad design". Without understanding your code it's quite hard to say for sure.
It is entirely possible that your problem is none of the ones I've mentioned here, but most likely it is one of these. Hopefully this asnwer is helpful to identify the cause.
Read CPU Caches and Why You Care to understand why a naive port of an algorithm from one thread to multiple threads will more often than not result in greatly reduced performance and negative scalability. Algorithms that are specififcally designed for parallelism take care of overactive interlocked operations, false sharing and other causes of cache pollution.
Here are a few things you might wanna look into.
1°) Do you enter any critical section (locks, semaphores, etc.) between your worker thread and your main thread? (this should be the case if your queries modify the graph). If so, that could be one of the sources of the multithreading overhead : threads competing for a lock usually degrades performances.
2°) You're using a 24 cores machines, which I assume would be NUMA (Non-Uniform Memory Access). Since you set the threads affinities during your tests, you should pay close attention to the memory topology of your hardware. Looking at the files in /sys/devices/system/cpu/cpuX/ can help you with that (beware that cpu0 and cpu1 aren't necessarily close together, and thus does not necessarily share memory). Threads heavily using memory should use local memory (allocated in the same NUMA node as the core they're executing on).
3°) You are heavily using disk I/O. Which kind of I/O is that? if every thread perform every time some synchronous I/O, you might wanna consider asynchronous system calls, so that the OS stays in charge of scheduling those requests to the disk.
4°) Some caches issues have already been mentionned in other answers. From experience, false sharing can hurt performances as much as you're observing. My last recommendation (which should have been my first) is to use a profiler tool, such as Linux Perf, or OProfile. With such performance degradation you're experiencing, the cause will certainly appear quite clearly.
The other answers have all addressed the general guidelines that can cause your symptoms. I will give my own, hopefully not excessively redundant version. Then I will talk a bit about how you can get to the bottom of the problem with everything discussed in mind.
In general, there's a few reasons you'd expect multiple threads to perform better:
A piece of work is dependent on some resources (disk, memory, cache, etc.) while other pieces can proceed independently of these resources or said workload.
You have multiple CPU cores that can process your workload in parallel.
The main reasons, enumerated above, you'd expect multiple threads to perform less well are all based on resource contention:
Disk contention: already explained in detail and can be a possible issue, especially if you are writing small buffers at a time instead of batching
CPU time contention if the threads are scheduled onto the same core: probably not your issue if you're setting affinity. However, you should still double check
Cache thrashing: similarly probably not your problem if you have affinity, though this can be very expensive if it is your problem.
Shared memory: again talked about in detail and doesn't seem to be your issue, but it wouldn't hurt to audit the code to check it out.
NUMA: again talked about. If your worker thread is pinned to a different core, you will want to check whether the work it needs to access is local to the main core.
Ok so far not much new. It can be any or none of the above. The question is, for your case, how can you detect where the extra time is coming from. There's a few strategies:
Audit the code and look for obvious areas. Don't spend too much time doing this as it's generally unfruitful if you wrote the program to begin with.
Refactor the single threaded code and the multi-threaded code to isolate one process() function, then profile at key checkpoints to try to account for the difference. Then narrow it down.
Refactor the resource access into batches, then profile each batch on both the control and the experiment to account for the difference. Not only will this tell you which areas (disk access vs memory access vs spending time in some tight loop) you need to focus your efforts on, doing this refactor might even improve your running time overall. Example:
First copy the graph structure to thread-local memory (perform a straight-up copy in the single-threaded case)
Then perform the query
Then setup an asynchronous write to disk
Try to find a minimally reproducible workload with the same symptoms. This means changing your algorithm to do a subset of what it already does.
Make sure there's no other noise in the system that could've caused the difference (if some other user is running a similar system on the work core).
My own intuition for your case:
Your graph structure is not NUMA friendly for your worker core.
The kernel can actually scheduled your worker thread off the affinity core. This can happen if you don't have isolcpu on for the core you're pinning to.
I can't tell you what's wrong with your program because you haven't shared enough of it to do a detailed analysis.
What I can tell you is if this was my problem the first thing I would try is to run two profiler sessions on my application, one on the single threaded version and another on the dual thread configuration. The profiler report should give you a pretty good idea of where the extra time is going. Note that you may not need to profile the entire application run, depending on the problem the time difference may become obvious after you profile for a few seconds or minutes.
As far as profiler choices for Linux you may want to consider oprofile or as a second choice gprof.
If you find you need help interpreting the profiler output feel free to add that to your question.
It can be a right pain in the rear to track down why threads aren't working as planned. One can do so analytically, or one can use tool to show what's going on. I've had very good mileage out of ftrace, Linux's clone of Solaris's dtrace (which in turn is based on what VxWorks, Greenhill's Integrity OS and Mercury Computer Systems Inc have been doing for a looong time.)
In particular I found this page very useful: http://www.omappedia.com/wiki/Installing_and_Using_Ftrace, particularly this and this section. Don't worry about it being an OMAP orientated website; I've used it on X86 Linuxes just fine (though you may have to build a kernel to include it). Also remember that the GTKWave viewer is primarily intended for looking at log traces from VHDL developments, which is why it looks 'odd'. It's just that someone realised that it would be a usable viewer for sched_switch data too, and that saved them writing one.
Using the sched_switch tracer you can see when (but not necessarily why) your threads are running, and that might be enough to give you a clue. The 'why' can be revealed by careful examination of some of the other tracers.
If you are getting slowdown from using 1 thread, it is likely due to overhead from using thread safe library functions, or from thread setup. Creating a thread for each job will cause significant overhead, but probably not as much as you refer to.
In other words, it is probably some overhead from some thread safe library function.
The best thing to do, is to profile your code to find out where time is spent. If it is in a library call, try to find a replacement library or implement it yourself. If the bottleneck is thread creation/destruction try reusing threads, for instance using OpenMP tasks or std::async in C++11.
Some libraries are really nasty wrt thread safe overhead. For instance, many rand() implementations use a global lock, rather than using thread local prgn's. Such locking overhead is much larger than generating a number, and is hard to track without a profiler.
The slowdown could also stem from small changes you have made, for instance declaring variables volatile, which generally should not be necessary.
I suspect you're running on a machine with one single-core processor. This problem is not parallelizable on that kind of system. Your code is constantly using the processor, which has a fixed number of cycles to offer to it. It actually runs more slowly because the additional thread adds expensive context switching to the problem.
The only kinds of problems that parallelize well on a single-processor machine are those that allow one path of execution to run while another is blocked waiting for I/O, and situations (such as keeping a responsive GUI) where allowing one thread to get some processor time is more important than executing your code as quickly as possible.
If you only want to run many independent instances of your algorithm can you just submit multiple jobs (with different parameters, can be handled by a single script) to your cluster? That would eliminate the need to profile and debug your multithreaded program. I don't have much experience with multithreaded programming but if you use MPI or OpenMP then you'd have to write less code for the book keeping too. For example, if some common initialization routine is needed and the processes can run independently thereafter you can just do that by initializing in one thread and doing a broadcast. No need for maintaining locks and such.
I've been looking into learning Erlang/OTP, and as a result, have been reading (okay, skimming) about the actor model.
From what I understand, the actor model is simply a set of functions (run within lightweight threads called "processes" in Erlang/OTP), which communicate with each other only via message passing.
This seems fairly trivial to implement in C++, or any other language:
class BaseActor {
std::queue<BaseMessage*> messages;
CriticalSection messagecs;
BaseMessage* Pop();
public:
void Push(BaseMessage* message)
{
auto scopedlock = messagecs.AquireScopedLock();
messagecs.push(message);
}
virtual void ActorFn() = 0;
virtual ~BaseActor() {} = 0;
}
With each of your processes being an instance of a derived BaseActor. Actors communicate with each other only via message-passing. (namely, pushing). Actors register themselves with a central map on initialization which allows other actors to find them, and allows a central function to run through them.
Now, I understand I'm missing, or rather, glossing over one important issue here, namely:
lack of yielding means a single Actor can unfairly consume excessive time. But are cross-platform coroutines the primary thing that makes this hard in C++? (Windows for instance has fibers.)
Is there anything else I'm missing, though, or is the model really this obvious?
The C++ code does not deal with fairness, isolation, fault detection or distribution which are all things which Erlang brings as part of its actor model.
No actor is allowed to starve any other actor (fairness)
If one actor crashes, it should only affect that actor (isolation)
If one actor crashes, other actors should be able to detect and react to that crash (fault detection)
Actors should be able to communicate over a network as if they were on the same machine (distribution)
Also the beam SMP emulator brings JIT scheduling of the actors, moving them to the core which is at the moment the one with least utilization and also hibernates the threads on certain cores if they are no longer needed.
In addition all the libraries and tools written in Erlang can assume that this is the way the world works and be designed accordingly.
These things are not impossible to do in C++, but they get increasingly hard if you add the fact that Erlang works on almost all of the major hw and os configurations.
edit: Just found a description by Ulf Wiger about what he sees erlang style concurrency as.
I don't like to quote myself, but from Virding's First Rule of Programming
Any sufficiently complicated concurrent program in another language contains an ad hoc informally-specified bug-ridden slow implementation of half of Erlang.
With respect to Greenspun. Joe (Armstrong) has a similar rule.
The problem is not to implement actors, that's not that difficult. The problem is to get everything working together: processes, communication, garbage collection, language primitives, error handling, etc ... For example using OS threads scales badly so you need to do it yourself. It would be like trying to "sell" an OO language where you can only have 1k objects and they are heavy to create and use. From our point of view concurrency is the basic abstraction for structuring applications.
Getting carried away so I will stop here.
This is actually an excellent question, and has received excellent answers that perhaps are yet unconvincing.
To add shade and emphasis to the other great answers already here, consider what Erlang takes away (compared to traditional general purpose languages such as C/C++) in order to achieve fault-tolerance and uptime.
First, it takes away locks. Joe Armstrong's book lays out this thought experiment: suppose your process acquires a lock and then immediately crashes (a memory glitch causes the process to crash, or the power fails to part of the system). The next time a process waits for that same lock, the system has just deadlocked. This could be an obvious lock, as in the AquireScopedLock() call in the sample code; or it could be an implicit lock acquired on your behalf by a memory manager, say when calling malloc() or free().
In any case, your process crash has now halted the entire system from making progress. Fini. End of story. Your system is dead. Unless you can guarantee that every library you use in C/C++ never calls malloc and never acquires a lock, your system is not fault tolerant. Erlang systems can and do kill processes at will when under heavy load in order make progress, so at scale your Erlang processes must be killable (at any single point of execution) in order to maintain throughput.
There is a partial workaround: using leases everywhere instead of locks, but you have no guarantee that all the libraries you utilize also do this. And the logic and reasoning about correctness gets really hairy quickly. Moreover leases recover slowly (after the timeout expires), so your entire system just got really slow in the face of failure.
Second, Erlang takes away static typing, which in turn enables hot code swapping and running two versions of the same code simultaneously. This means you can upgrade your code at runtime without stopping the system. This is how systems stay up for nine 9's or 32 msec of downtime/year. They are simply upgraded in place. Your C++ functions will have to be manually re-linked in order to be upgraded, and running two versions at the same time is not supported. Code upgrades require system downtime, and if you have a large cluster that cannot run more than one version of code at once, you'll need to take the entire cluster down at once. Ouch. And in the telecom world, not tolerable.
In addition Erlang takes away shared memory and shared shared garbage collection; each light weight process is garbage collected independently. This is a simple extension of the first point, but emphasizes that for true fault tolerance you need processes that are not interlocked in terms of dependencies. It means your GC pauses compared to java are tolerable (small instead of pausing a half-hour for a 8GB GC to complete) for big systems.
There are actual actor libraries for C++:
http://actor-framework.org/
http://www.theron-library.com/
And a list of some libraries for other languages.
It is a lot less about the actor model and a lot more about how hard it is to properly write something analogous to OTP in C++. Also, different operating systems provide radically different debugging and system tooling, and Erlang's VM and several language constructs support a uniform way of figuring out just what all those processes are up to which would be very hard to do in a uniform way (or maybe do at all) across several platforms. (It is important to remember that Erlang/OTP predates the current buzz over the term "actor model", so in some cases these sort of discussions are comparing apples and pterodactyls; great ideas are prone to independent invention.)
All this means that while you certainly can write an "actor model" suite of programs in another language (I know, I have done this for a long time in Python, C and Guile without realizing it before I encountered Erlang, including a form of monitors and links, and before I'd ever heard the term "actor model"), understanding how the processes your code actually spawns and what is happening amongst them is extremely difficult. Erlang enforces rules that an OS simply can't without major kernel overhauls -- kernel overhauls that would probably not be beneficial overall. These rules manifest themselves as both general restrictions on the programmer (which can always be gotten around if you really need to) and basic promises the system guarantees for the programmer (which can be deliberately broken if you really need to also).
For example, it enforces that two processes cannot share state to protect you from side effects. This does not mean that every function must be "pure" in the sense that everything is referentially transparent (obviously not, though making as much of your program referentially transparent as practical is a clear design goal of most Erlang projects), but rather that two processes aren't constantly creating race conditions related to shared state or contention. (This is more what "side effects" means in the context of Erlang, by the way; knowing that may help you decipher some of the discussion questioning whether Erlang is "really functional or not" when compared with Haskell or toy "pure" languages.)
On the other hand, the Erlang runtime guarantees delivery of messages. This is something sorely missed in an environment where you must communicate purely over unmanaged ports, pipes, shared memory and common files which the OS kernel is the only one managing (and OS kernel management of these resources is necessarily extremely minimal compared to what the Erlang runtime provides). This doesn't meant that Erlang guarantees RPC (anyway, message passing is not RPC, nor is it method invocation!), it doesn't promise that your message is addressed correctly, and it doesn't promise that a process you're trying to send a message to exists or is alive, either. It just guarantees delivery if the thing your sending to happens to be valid at that moment.
Built on this promise is the promise that monitors and links are accurate. And based on that the Erlang runtime makes the entire concept of "network cluster" sort of melt away once you grasp what is going on with the system (and how to use erl_connect...). This permits you to hop over a set of tricky concurrency cases already, which gives one a big head start on coding for the successful case instead of getting mired in the swamp of defensive techniques required for naked concurrent programming.
So its not really about needing Erlang, the language, its about the runtime and OTP already existing, being expressed in a rather clean way, and implementing anything close to it in another language being extremely hard. OTP is just a hard act to follow. In the same vein, we don't really need C++, either, we could just stick to raw binary input, Brainfuck and consider Assembler our high level language. We also don't need trains or ships, as we all know how to walk and swim.
All that said, the VM's bytecode is well documented, and a number of alternative languages have emerged that compile to it or work with the Erlang runtime. If we break the question into a language/syntax part ("Do I have to understand Moon Runes to do concurrency?") and a platform part ("Is OTP the most mature way to do concurrency, and will it guide me around the trickiest, most common pitfalls to be found in a concurrent, distributed environment?") then the answer is ("no", "yes").
Casablanca is another new kid on the actor model block. A typical asynchronous accept looks like this:
PID replyTo;
NameQuery request;
accept_request().then([=](std::tuple<NameQuery,PID> request)
{
if (std::get<0>(request) == FirstName)
std::get<1>(request).send("Niklas");
else
std::get<1>(request).send("Gustafsson");
}
(Personally, I find that CAF does a better job at hiding the pattern matching behind a nice interface.)
As software gets more and more concurrent, how do you handle testing the core behaviour of the type with your unit tests (not the parallel behaviour, just the core behaviour)?
In the good old days, you had a type, you called it, and you checked either what it returned and/or what other things it called.
Nowadays, you call a method and the actual work gets scheduled to run on the next available thread; you don't know when it'll actually start and call the other things - and what's more, those other things could be concurrent too.
How do you deal with this? Do you abstract/inject the concurrent scheduler (e.g. abstract the Task Parallel Library and provide a fake/mock in the unit tests)?
What resources have you come across that helped you?
Edit
I've edited the question to emphasise testing the normal behaviour of the type (ignoring whatever parallel mechanism is used to take advantage of multi-core, e.g. the TPL)
Disclaimer: I work for Corensic, a small startup in Seattle. We've got a tool called Jinx that is designed to detect concurrency errors in your code. It's free for now while we're in Beta, so you might want to check it out. ( http://www.corensic.com/ )
In a nutshell, Jinx is a very thin hypervisor that, when activated, slips in between the processor and operating system. Jinx then intelligently takes slices of execution and runs simulations of various thread timings to look for bugs. When we find a particular thread timing that will cause a bug to happen, we make that timing "reality" on your machine (e.g., if you're using Visual Studio, the debugger will stop at that point). We then point out the area in your code where the bug was caused. There are no false positives with Jinx. When it detects a bug, it's definitely a bug.
Jinx works on Linux and Windows, and in both native and managed code. It is language and application platform agnostic and can work with all your existing tools.
If you check it out, please send us feedback on what works and doesn't work. We've been running Jinx on some big open source projects and already are seeing situations where Jinx can find bugs 50-100 times faster than simply stress testing code.
I recommend picking up a copy of Growing Object Oriented Software by Freeman and Pryce. The last couple of chapters are very enlightening and deal with this specific topic. It also introduces some terminology which helps in pinning down the notation for discussion.
To summarize ....
Their core idea is to split the functionality and concurrent/synchronization aspects.
First test-drive the functional part in a single synchronous thread like a normal object.
Once you have the functional part pinned down. You can move on to the concurrent aspect. To do that, you'd have to think and come up with "observable invariants w.r.t. concurrency" for your object, e.g. the count should be equal to the times the method is called. Once you have identified the invariants, you can write stress tests that run multiple threads et.all to try and break your invariants. The stress tests assert your invariants.
Finally as an added defence, run tools or static analysis to find bugs.
For passive objects, i.e. code that'd be called from clients on different threads: your test needs to mimic clients by starting its own threads. You would then need to choose between a notification-listening or sampling/polling approach to synchronize your tests with the SUT.
You could either block till you receive an expected notification
Poll certain observable side-effects with a reasonable timeout.
The field of Unit testing for race conditions and deadlocks is relativly new and lacks good tools.
I know of two such tools both in early alpha/beta stages:
Microsoft's Chess
Typemock Racer
ANother option is to try and write a "stress test" that would cause deadlocks/race condtions to surface, create multiople instances/threads and run them side by side. The downside of this approch is that if the test fail it would be very hard to reproduce it. I suggest using logs both in the test and production code so that you'll be able to understand what happened.
A technique I've found useful is to run tests within a tool that detects race conditions like Intel Parallel Inspector. The test runs much slower than normal, because dependencies on timing have to be checked, but a single run can find bugs that otherwise would require millions of repeated ordinary runs.
I've found this very useful when converting existing systems for fine-grained parallelism via multi-core.
Unit tests really should not test concurrency/asynchronous behaviour, you should use mocks there and verify that the mocks receive the expected input.
For integration tests I just explicitly call the background task, then check the expectations after that.
In Cucumber it looks like this:
When I press "Register"
And the email sending script is run
Then I should have an email
Given that your TPL will have its own separate unit test you don't need to verify that.
Given that I write two tests for each module:
1) A single threaded unit test that uses some environment variable or #define to turn of the TPL so that I can test my module for functional correctness.
2) A stress test that runs the module in its threaded deployable mode. This test attempts to find concurrency issues and should use lots of random data.
The second test often includes many modules and so is probably more of an integration/system test.
My company's main product is a large monolithic C++ application, used for scientific data processing and visualisation. Its codebase goes back maybe 12 or 13 years, and while we have put work into upgrading and maintaining it (use of STL and Boost - when I joined most containers were custom, for example - fully upgraded to Unicode and the 2010 VCL, etc) there's one remaining, very significant problem: it's fully singlethreaded. Given it's a data processing and visualisation program, this is becoming more and more of a handicap.
I'm both a developer and the project manager for the next release where we want to tackle this, and this is going to be a difficult job in both areas. I'm seeking concrete, practical, and architectural advice on how to tackle the problem.
The program's data flow might go something like this:
a window needs to draw data
In the paint method, it will call a GetData method, often hundreds of times for hundreds of bits of data in one paint operation
This will go and calculate or read from file or whatever else is required (often quite a complex data flow - think of this as data flowing through a complex graph, each node of which performs operations)
Ie, the paint message handler will block while processing is done, and if the data hasn't already been calculated and cached, this can be a long time. Sometimes this is minutes. Similar paths occur for other parts of the program that perform lengthy processing operations - the program is unresponsive for the entire time, sometimes hours.
I'm seeking advice on how to approach changing this. Practical ideas. Perhaps things like:
design patterns for asynchronously requesting data?
storing large collections of objects such that threads can read and write safely?
handling invalidation of data sets while something is trying to read it?
are there patterns and techniques for this sort of problem?
what should I be asking that I haven't thought of?
I haven't done any multithreaded programming since my Uni days a few years ago, and I think the rest of my team is in a similar position. What I knew was academic, not practical, and is nowhere near enough to have confidence approaching this.
The ultimate objective is to have a fully responsive program, where all calculations and data generation is done in other threads and the UI is always responsive. We might not get there in a single development cycle :)
Edit: I thought I should add a couple more details about the app:
It's a 32-bit desktop application for Windows. Each copy is licensed. We plan to keep it a desktop, locally-running app
We use Embarcadero (formerly Borland) C++ Builder 2010 for development. This affects the parallel libraries we can use, since most seem (?) to be written for GCC or MSVC only. Luckily they're actively developing it and its C++ standards support is much better than it used to be. The compiler supports these Boost components.
Its architecture is not as clean as it should be and components are often too tightly coupled. This is another problem :)
Edit #2: Thanks for the replies so far!
I'm surprised so many people have recommended a multi-process architecture (it's the top-voted answer at the moment), not multithreading. My impression is that's a very Unix-ish program structure, and I don't know anything about how it's designed or works. Are there good resources available about it, on Windows? Is it really that common on Windows?
In terms of concrete approaches to some of the multithreading suggestions, are there design patterns for asynchronous request and consuming of data, or threadaware or asynchronous MVP systems, or how to design a task-oriented system, or articles and books and post-release deconstructions illustrating things that work and things that don't work? We can develop all this architecture ourselves, of course, but it's good to work from what others have done before and know what mistakes and pitfalls to avoid.
One aspect that isn't touched on in any answers is project managing this. My impression is estimating how long this will take and keeping good control of the project when doing something as uncertain as this may be hard. That's one reason I'm after recipes or practical coding advice, I guess, to guide and restrict coding direction as much as possible.
I haven't yet marked an answer for this question - this is not because of the quality of the answers, which is great (and thankyou) but simply that because of the scope of this I'm hoping for more answers or discussion. Thankyou to those who have already replied!
You have a big challenge ahead of you. I had a similar challenge ahead of me -- 15 year old monolithic single threaded code base, not taking advantage of multicore, etc. We expended a great deal of effort in trying to find a design and solution that was workable and would work.
Bad news first. It will be somewhere between impractical and impossible to make your single-threaded app multithreaded. A single threaded app relies on it's singlethreaded-ness is ways both subtle and gross. One example is if the computation portion requires input from the GUI portion. The GUI must run in the main thread. If you try to get this data directly from the computation engine, you will likely run in to deadlock and race conditions that will require major redesigns to fix. Many of these reliances will not crop up during the design phase, or even during the development phase, but only after a release build is put in a harsh environment.
More bad news. Programming multithreaded applications is exceptionally hard. It might seem fairly straightforward to just lock stuff and do what you have to do, but it is not. First of all if you lock everything in sight you end up serializing your application, negating every benefit of mutithreading in the first place while still adding in all the complexity. Even if you get beyond this, writing a defect-free MP application is hard enough, but writing a highly-performant MP application is that much more difficult. You could learn on the job in a kind of baptismal by fire. But if you are doing this with production code, especially legacy production code, you put your buisness at risk.
Now the good news. You do have options that don't involve refactoring your whole app and will give you most of what you seek. One option in particular is easy to implement (in relative terms), and much less prone to defects than making your app fully MP.
You could instantiate multiple copies of your application. Make one of them visible, and all the others invisible. Use the visible application as the presentation layer, but don't do the computational work there. Instead, send messages (perhaps via sockets) to the invisible copies of your application which do the work and send the results back to the presentation layer.
This might seem like a hack. And maybe it is. But it will get you what you need without putting the stability and performance of your system at such great risk. Plus there are hidden benefits. One is that the invisible engine copies of your app will have access to their own virtual memory space, making it easier to leverage all the resources of the system. It also scales nicely. If you are running on a 2-core box, you could spin off 2 copies of your engine. 32 cores? 32 copies. You get the idea.
So, there's a hint in your description of the algorithm as to how to proceed:
often quite a complex data flow - think of this as data flowing through a complex graph, each node of which performs operations
I'd look into making that data-flow graph be literally the structure that does the work. The links in the graph can be thread-safe queues, the algorithms at each node can stay pretty much unchanged, except wrapped in a thread that picks up work items from a queue and deposits results on one. You could go a step further and use sockets and processes rather than queues and threads; this will let you spread across multiple machines if there is a performance benefit in doing this.
Then your paint and other GUI methods need split in two: one half to queue the work, and the other half to draw or use the results as they come out of the pipeline.
This may not be practical if the app presumes that data is global. But if it is well contained in classes, as your description suggests it may be, then this could be the simplest way to get it parallelised.
Don't attempt to multithread everything in the old app. Multithreading for the sake of saying it's multithreaded is a waste of time and money. You're building an app that does something, not a monument to yourself.
Profile and study your execution flows to figure out where the app spends most of its time. A profiler is a great tool for this, but so is just stepping through the code in the debugger. You find the most interesting things in random walks.
Decouple the UI from long-running computations. Use cross-thread communications techniques to send updates to the UI from the computation thread.
As a side-effect of #3: think carefully about reentrancy: now that the compute is running in the background and the user can smurf around in the UI, what things in the UI should be disabled to prevent conflicts with the background operation? Allowing the user to delete a dataset while a computation is running on that data is probably a bad idea. (Mitigation: computation makes a local snapshot of the data) Does it make sense for the user to spool up multiple compute operations concurrently? If handled well, this could be a new feature and help rationalize the app rework effort. If ignored, it will be a disaster.
Identify specific operations that are candidates to be shoved into a background thread. The ideal candidate is usually a single function or class that does a lot of work (requires a "lot of time" to complete - more than a few seconds) with well defined inputs and outputs, that makes use of no global resources, and does not touch the UI directly. Evaluate and prioritize candidates based on how much work would be required to retrofit to this ideal.
In terms of project management, take things one step at a time. If you have multiple operations that are strong candidates to be moved to a background thread, and they have no interaction with each other, these might be implemented in parallel by multiple developers. However, it would be a good exercise to have everybody participate in one conversion first so that everyone understands what to look for and to establish your patterns for UI interaction, etc. Hold an extended whiteboard meeting to discuss the design and process of extracting the one function into a background thread. Go implement that (together or dole out pieces to individuals), then reconvene to put it all together and discuss discoveries and pain points.
Multithreading is a headache and requires more careful thought than straight up coding, but splitting the app into multiple processes creates far more headaches, IMO. Threading support and available primitives are good in Windows, perhaps better than some other platforms. Use them.
In general, don't do any more than what is needed. It's easy to severely over implement and over complicate an issue by throwing more patterns and standard libraries at it.
If nobody on your team has done multithreading work before, budget time to make an expert or funds to hire one as a consultant.
The main thing you have to do is to disconnect your UI from your data set. I'd suggest that the way to do that is to put a layer in between.
You will need to design a data structure of data cooked-for-display. This will most likely contain copies of some of your back-end data, but "cooked" to be easy to draw from. The key idea here is that this is quick and easy to paint from. You may even have this data structure contain calculated screen positions of bits of data so that it's quick to draw from.
Whenever you get a WM_PAINT message you should get the most recent complete version of this structure and draw from it. If you do this properly, you should be able to handle multiple WM_PAINT messages per second because the paint code never refers to your back end data at all. It's just spinning through the cooked structure. The idea here is that its better to paint stale data quickly than to hang your UI.
Meanwhile...
You should have 2 complete copies of this cooked-for-display structure. One is what the WM_PAINT message looks at. (call it cfd_A) The other is what you hand to your CookDataForDisplay() function. (call it cfd_B). Your CookDataForDisplay() function runs in a separate thread, and works on building/updating cfd_B in the background. This function can take as long as it wants because it isn't interacting with the display in any way. Once the call returns cfd_B will be the most up-to-date version of the structure.
Now swap cfd_A and cfd_B and InvalidateRect on your application window.
A simplistic way to do this is to have your cooked-for-display structure be a bitmap, and that might be a good way to go to get the ball rolling, but I'm sure with a bit of thought you can do a much better job with a more sophisticated structure.
So, referring back to your example.
In the paint method, it will call a GetData method, often hundreds of times for hundreds of bits of data in one paint operation
This is now 2 threads, the paint method refers to cfd_A and runs on the UI thread. Meanwhile cfd_B is being built by a background thread using GetData calls.
The quick-and-dirty way to do this is
Take your current WM_PAINT code, stick it into a function called PaintIntoBitmap().
Create a bitmap and a Memory DC, this is cfd_B.
Create a thread and pass it cfd_B and have it call PaintIntoBitmap()
When this thread completes, swap cfd_B and cfd_A
Now your new WM_PAINT method just takes the pre-rendered bitmap in cfd_A and draws it to the screen. Your UI is now disconnnected from your backend GetData() function.
Now the real work begins, because the quick-and-dirty way doesn't handle window resizing very well. You can go from there to refine what your cfd_A and cfd_B structures are a little at a time until you reach a point where you are satisfied with the result.
You might just start out breaking the the UI and the work task into separate threads.
In your paint method instead of calling getData() directly, it puts the request in a thread-safe queue. getData() is run in another thread that reads its data from the queue. When the getData thread is done, it signals the main thread to redraw the visualisation area with its result data using thread syncronization to pass the data.
While all this is going on you of course have a progress bar saying reticulating splines so the user knows something is going on.
This would keep your UI snappy without the significant pain of multithreading your work routines (which can be akin to a total rewrite)
It sounds like you have several different issues that parallelism can address, but in different ways.
Performance increases through utilizing multicore CPU Architecutres
You're not taking advantage of the multi-core CPU architetures that are becoming so common. Parallelization allow you to divide work amongst multiple cores. You can write that code through standard C++ divide and conquer techniques using a "functional" style of programming where you pass work to separate threads at the divide stage. Google's MapReduce pattern is an example of that technique. Intel has the new CILK library to give you C++ compiler support for such techniques.
Greater GUI responsiveness through asynchronous document-view
By separating the GUI operations from the document operations and placing them on different threads, you can increase the apparent responsiveness of your application. The standard Model-View-Controller or Model-View-Presenter design patterns are a good place to start. You need to parallelize them by having the model inform the view of updates rather than have the view provide the thread on which the document computes itself. The View would call a method on the model asking it to compute a particular view of the data, and the model would inform the presenter/controller as information is changed or new data becomes available, which would get passed to the view to update itself.
Opportunistic caching and pre-calculation
It sounds like your application has a fixed base of data, but many possible compute-intensive views on the data. If you did a statistical analysis on which views were most commonly requested in what situations, you could create background worker threads to pre-calculate the likely-requested values. It may be useful to put these operations on low-priority threads so that they don't interfere with the main application processing.
Obviously, you'll need to use mutexes (or critical sections), events, and probably semaphores to implement this. You may find some of the new synchronization objects in Vista useful, like the slim reader-writer lock, condition variables, or the new thread pool API. See Joe Duffy's book on concurrency for how to use these basic techniques.
There is something that no-one has talked about yet, but which is quite interesting.
It's called futures. A future is the promise of a result... let's see with an example.
future<int> leftVal = computeLeftValue(treeNode); // [1]
int rightVal = computeRightValue(treeNode); // [2]
result = leftVal + rightVal; // [3]
It's pretty simple:
You spin off a thread that starts computing leftVal, taking it from a pool for example to avoid the initialization problem.
While leftVal is being computed, you compute rightVal.
You add the two, this may block if leftVal is not computed yet and wait for the computation to end.
The great benefit here is that it's straightforward: each time you have one computation followed by another that is independent and you then join the result, you can use this pattern.
See Herb Sutter's article on futures, they will be available in the upcoming C++0x but there are already libraries available today even if the syntax is perhaps not as pretty as I would make you believe ;)
If it was my development dollars I was spending, I would start with the big picture:
What do I hope to accomplish, and how much will I spend to accomplish this, and how will I be further ahead? (If the answer to this is, my app will run 10% better on quadcore PCs, and I could have achieved the same result by spending $1000 more per customer PC , and spending $100,000 less this year on R&D, then, I would skip the whole effort).
Why am I doing multi-threaded instead of massively parallel distributed? Do I really think threads are better than processes? Multi-core systems also run distributed apps pretty well. And there are some advantages to message-passing process based systems that go beyond the benefits (and the costs!) of threading. Should I consider a process-based approach? SHould I consider a background running entirely as a service, and a foreground GUI? Since my product is node-locked and licensed, I think services would suit me (vendor) quite well. Also, separating stuff into two processes (background service and foreground) just might force the kind of rewrite and rearchitecting to occur that I might not be forced to do, if I was to just add threading into my mix.
This is just to get you thinking: What if you were to rewrite it as a service (background app) and a GUI, because that would actually be easier than adding threading, without also adding crashes, deadlocks, and race conditions?
Consider the idea that for your needs, perhaps threading is evil. Develop your religion, and stick with that. Unless you have a real good reason to go the other way. For many years, I religiously avoided threading. Because one thread per process is good enough for me.
I don't see any really solid reasons in your list why you need threading, except ones that could be more inexpensively solved by more expensive target computer hardware. If your app is "too slow" adding in threads might not even speed it up.
I use threads for background serial communications, but I would not consider threading merely for computationally heavy applications, unless my algorithms were so inherently parallel as to make the benefits clear, and the drawbacks minimal.
I wonder if the "design" problems that this C++Builder app has are like my Delphi "RAD Spaghetti" application disease. I have found that a wholesale refactor/rewrite (over a year per major app that I have done this to), was a minimum amount of time for me to get a handle on application "accidental complexity". And that was without throwing a "threads where possible" idea. I tend to write my apps with threads for serial communication and network socket handling, only. And maybe the odd "worker-thread-queue".
If there is a place in your app you can add ONE thread, to test the waters, I would look for the main "work queue" and I would create an experimental version control branch, and I would learn about how my code works by breaking it in the experimental branch. Add that thread. And see where you spend your first day of debugging. Then I might just abandon that branch and go back to my trunk until the pain in my temporal lobe subsides.
Warren
Here's what I would do...
I would start by profiling your and seeing:
1) what is slow and what the hot paths are
2) which calls are reentrant or deeply nested
you can use 1) to determine where the opportunity is for speedups and where to start looking for parallelization.
you can use 2) to find out where the shared state is likely to be and get a deeper sense of how much things are tangled up.
I would use a good system profiler and a good sampling profiler (like the windows perforamnce toolkit or the concurrency views of the profiler in Visual Studio 2010 Beta2 - these are both 'free' right now).
Then I would figure out what the goal is and how to separate things gradually to a cleaner design that is more responsive (moving work off the UI thread) and more performant (parallelizing computationally intensive portions). I would focus on the highest priority and most noticable items first.
If you don't have a good refactoring tool like VisualAssist, invest in one - it's worth it. If you're not familiar with Michael Feathers or Kent Beck's refactoring books, consider borrowing them. I would ensure my refactorings are well covered by unit tests.
You can't move to VS (I would recommend the products I work on the Asynchronous Agents Library & Parallel Pattern Library, you can also use TBB or OpenMP).
In boost, I would look carefully at boost::thread, the asio library and the signals library.
I would ask for help / guidance / a listening ear when I got stuck.
-Rick
You can also look at this article from Herb Sutter You have a mass of existing code and want to add concurrency. Where do you start?
Well, I think you're expecting a lot based on your comments here. You're not going to go from minutes to milliseconds by multithreading. The most you can hope for is the current amount of time divided by the number of cores. That being said, you're in a bit of luck with C++. I've written high performance multiprocessor scientific apps, and what you want to look for is the most embarrassingly parallel loop you can find. In my scientific code, the heaviest piece is calculating somewhere between 100 and 1000 data points. However, all of the data points can be calculated independently of the others. You can then split the loop using openmp. This is the easiest and most efficient way to go. If you're compiler doesn't support openmp, then you will have a very hard time porting existing code. With openmp (if you're lucky), you may only have to add a couple of #pragmas to get 4-8x the performance. Here's an example StochFit
I hope this will help you in understanding and converting your monolithic single threaded app to multi thread easily. Sorry it is for another programming language but never the less the principles explained are the same all over.
http://www.freevbcode.com/ShowCode.Asp?ID=1287
Hope this helps.
The first thing you must do is to separate your GUI from your data, the second is to create a multithreaded class.
STEP 1 - Responsive GUI
We can assume that the image you are producing is contained in the canvas of a TImage. You can put a simple TTimer in you form and you can write code like this:
if (CurrenData.LastUpdate>CurrentUpdate)
{
Image1->Canvas->Draw(0,0,CurrenData.Bitmap);
CurrentUpdate=Now();
}
OK! I know! Is a little bit dirty, but it's fast and is simple.The point is that:
You need an Object that is created in the main thread
The object is copied in the Form you need, only when is needed and in a safe way (ok, a better protection for the Bitmap may be is needed, but for semplicity...)
The object CurrentData is your actual project, single threaded, that produces an image
Now you have a fast and responsive GUI. If your algorithm as slow, the refresh is slow, but your user will never think that your program is freezed.
STEP 2 - Multithread
I suggest you to implement a class like the following:
SimpleThread.h
typedef void (__closure *TThreadFunction)(void* Data);
class TSimpleThread : public TThread
{
public:
TSimpleThread( TThreadFunction _Action,void* _Data = NULL, bool RunNow = true );
void AbortThread();
__property Terminated;
protected:
TThreadFunction ThreadFunction;
void* Data;
private:
virtual void __fastcall Execute() { ThreadFunction(Data); };
};
SimpleThread.c
TSimpleThread::TSimpleThread( TThreadFunction _Action,void* _Data, bool RunNow)
: TThread(true), // initialize suspended
ThreadFunction(_Action), Data(_Data)
{
FreeOnTerminate = false;
if (RunNow) Resume();
}
void TSimpleThread::AbortThread()
{
Suspend(); // Can't kill a running thread
Free(); // Kills thread
}
Let's explain. Now, in your simple threaded class you can create an object like this:
TSimpleThread *ST;
ST=new TSimpleThread( RefreshFunction,NULL,true);
ST->Resume();
Let's explain better: now, in your own monolithic class, you have created a thread. More: you bring a function (ie: RefreshFunction) in a separate thread. The scope of your funcion is the same, the class is the same, the execution is separate.
My number one suggestion, although it's very late (sorry for reviving old thread, it's interesting!) is seek out homogeneous transform loops where each iteration of the loop is mutating a completely independent piece of data from the other iterations.
Instead of thinking about how to turn this old codebase into an asynchronous one running all kinds of operations in parallel (which could be asking for all kinds of trouble from worse than single-threaded performance from poor locking patterns or exponentially worse, race conditions/deadlocks by trying to do this in hindsight to code you can't fully comprehend), stick to the sequential mindset for the overall application design for now but identify or extract simple, homogeneous transform loops. Don't go from intrusive broad design-level multithreading and then try to drill into details. Work from non-intrusive multithreading of fine implementation details and specific hotspots first.
What I mean by homogeneous loops is basically one that transforms data in a very straightforward way, like:
for each pixel in image:
make it brighter
That is very simple to reason about and you can safely parallelize this loop without any problems whatsoever using OMP or TBB or whatever and without getting tangled up in thread synchronization. It only takes one glance at this code to fully comprehend its side effects.
Try to find as many hotspots as you can which fit this type of simple homogeneous transform loop and if you have complex loops which update many different types of data with complex control flows that trigger complex side effects, then seek to refactor towards these homogeneous loops. Often a complex loop which causes 3 disparate side effects to 3 different types of data can be turned into 3 simple homogeneous loops which each trigger just one kind of side effect to one type of data with a simpler control flow. Doing multiple loops instead of one might seem a tad wasteful, but the loops become simpler, the homogeneity will often lead to more cache-friendly sequential memory access patterns vs. sporadic random-access patterns, and you then tend to find much more opportunities to safely parallelize (as well as vectorize) the code in a straightforward way.
First you have to thoroughly understand the side effects of any code you attempt to parallelize (and I mean thoroughly!!!), so seeking out these homogeneous loops gives you isolated areas of the codebase you can easily reason about in terms of the side effects to the point where you can confidently and safely parallelize those hotspots. It'll also improve the maintainability of the code by making it very easy to reason about the state changes going on in that particular piece of code. Save the dream of the uber multithreaded application running everything in parallel for later. For now, focus on identifying/extracting performance-critical, homogeneous loops with simple control flows and simple side effects. Those are your priority targets for parallelization with simple parallelized loops.
Now admittedly I somewhat dodged your questions, but most of them don't need apply if you do what I suggest, at least until you've kind of worked your way out to the point where you're thinking more about multithreading designs as opposed to simply parallelizing implementation details. And you might not even need to go that far to have a very competitive product in terms of performance. If you have beefy work to do in a single loop, you can devote the hardware resources to making that loop go faster instead of making many operations run simultaneously. If you have to resort to more async methods like if your hotspots are more I/O bound, seek an async/wait approach where you fire off an async task but do some things in the meantime and then wait on the async task(s) to complete. Even if that's not absolutely necessary, the idea is to section off isolated areas of your codebase where you can, with 100% confidence (or at least 99.9999999%) say that the multithreaded code is correct.
You don't ever want to gamble with race conditions. There's nothing more demoralizing than finding some obscure race condition that only occurs once in a full moon on some random user's machine while your entire QA team is unable to reproduce it, only to, 3 months later, run into it yourself except during that one time you ran a release build without debugging info available while you then toss and turn in your sleep knowing your codebase can flake out at any given moment but in ways that no one will ever be able to consistently reproduce. So take it easy with multithreading legacy codebases, at least for now, and stick to multithreading isolated but critical sections of the codebase where the side effects are dead simple to reason about. And test the crap out of it -- ideally apply a TDD approach where you write a test for the code you're going to multithread to ensure it gives the correct output after you finish... though race conditions are the types of things that easily fly under the radar of unit and integration testing, so again you absolutely need to be able to comprehend the entirety of the side effects that go on in a given piece of code before you attempt to multithread it. The best way to do that is to make the side effects as easy to comprehend as possible with the simplest control flows causing just one type of side effect for an entire loop.
It is hard to give you proper guidelines. But...
The easiest way out according to me is to convert your application to ActiveX EXE as COM has support for Threading, etc. built right into it your program will automatically become Multi Threading application. Of course you will have to make quite a few changes to your code. But this is the shortest and safest way to go.
I am not sure but probably RichClient Toolset lib may do the trick for you. On the site the author has written:
It also offers registration free Loading/Instancing-capabilities
for ActiveX-Dlls and new, easy to use Threading-approach,
which works with Named-Pipes under the
hood and works therefore also
cross-process.
Please check it out. Who knows it may be the right solution for your requirements.
As for Project management I think you can continue using what is provided in your choice IDE by integrating it with SVN through plugins.
I forgot to mention that we have completed an application for Share market that automatically trades (buys and sells based on lows and highs) into those scripts that are in user portfolio based on an algorithm that we have developed.
While developing this software we were facing the same kind of problem as you have illustrated here. To solve it we converted out application in ActiveX EXE and we converted all those parts that need to execute parallely into ActiveX DLLs. We have not used any third party libs for this!
HTH
I'm currently working on a wireless networking application in C++ and it's coming to a point where I'm going to want to multi-thread pieces of software under one process, rather than have them all in separate processes. Theoretically, I understand multi-threading, but I've yet to dive in practically.
What should every programmer know when writing multi-threaded code in C++?
I would focus on design the thing as much as partitioned as possible so you have the minimal amount of shared things across threads. If you make sure you don't have statics and other resources shared among threads (other than those that you would be sharing if you designed this with processes instead of threads) you would be fine.
Therefore, while yes, you have to have in mind concepts like locks, semaphores, etc, the best way to tackle this is to try to avoid them.
I am no expert at all in this subject. Just some rule of thumb:
Design for simplicity, bugs really are hard to find in concurrent code even in the simplest examples.
C++ offers you a very elegant paradigm to manage resources(mutex, semaphore,...): RAII. I observed that it is much easier to work with boost::thread than to work with POSIX threads.
Build your code as thread-safe. If you don't do so, your program could behave strangely
I am exactly in this situation: I wrote a library with a global lock (many threads, but only one running at a time in the library) and am refactoring it to support concurrency.
I have read books on the subject but what I learned stands in a few points:
think parallel: imagine a crowd passing through the code. What happens when a method is called while already in action ?
think shared: imagine many people trying to read and alter shared resources at the same time.
design: avoid the problems that points 1 and 2 can raise.
never think you can ignore edge cases, they will bite you hard.
Since you cannot proof-test a concurrent design (because thread execution interleaving is not reproducible), you have to ensure that your design is robust by carefully analyzing the code paths and documenting how the code is supposed to be used.
Once you understand how and where you should bottleneck your code, you can read the documentation on the tools used for this job:
Mutex (exclusive access to a resource)
Scoped Locks (good pattern to lock/unlock a Mutex)
Semaphores (passing information between threads)
ReadWrite Mutex (many readers, exclusive access on write)
Signals (how to 'kill' a thread or send it an interrupt signal, how to catch these)
Parallel design patterns: boss/worker, producer/consumer, etc (see schmidt)
platform specific tools: openMP, C blocks, etc
Good luck ! Concurrency is fun, just take your time...
You should read about locks, mutexes, semaphores and condition variables.
One word of advice, if your app has any form of UI make sure you always change it from the UI thread. Most UI toolkits/frameworks will crash (or behave unexpectedly) if you access them from a background thread. Usually they provide some form of dispatching method to execute some function in the UI thread.
Never assume that external APIs are threadsafe. If it is not explicitly stated in their docs, do not call them concurrently from multiple threads. Instead, limit your use of them to a single thread or use a mutex to prevent concurrent calls (this is rather similar to the aforementioned GUI libraries).
Next point is language-related. Remember, C++ has (currently) no well-defined approach to threading. The compiler/optimizer does not know if code might be called concurrently. The volatile keyword is useful to prevent certain optimizations (i.e. caching of memory fields in CPU registers) in multi-threaded contexts, but it is no synchronization mechanism.
I'd recommend boost for synchronization primitives. Don't mess with platform APIs. They make your code difficult to port because they have similar functionality on all major platforms, but slightly different detail behaviour. Boost solves these problems by exposing only common functionality to the user.
Furthermore, if there's even the smallest chance that a data structure could be written to by two threads at the same time, use a synchronization primitive to protect it. Even if you think it will only happen once in a million years.
One thing I've found very useful is to make the application configurable with regard to the actual number of threads it uses for various tasks. For example, if you have multiple threads accessing a database, make the number of those threads be configurable via a command line parameter. This is extremely handy when debugging - you can exclude threading issues by setting the number to 1, or force them by setting it to a high number. It's also very handy when working out what the optimal number of threads is.
Make sure you test your code in a single-cpu system and a multi-cpu system.
Based on the comments:-
Single socket, single core
Single socket, two cores
Single socket, more than two cores
Two sockets, single core each
Two sockets, combination of single, dual and multi core cpus
Mulitple sockets, combination of single, dual and multi core cpus
The limiting factor here is going to be cost. Ideally, concentrate on the types of system your code is going to run on.
In addition to the other things mentioned, you should learn about asynchronous message queues. They can elegantly solve the problems of data sharing and event handling. This approach works well when you have concurrent state machines that need to communicate with each other.
I'm not aware of any message passing frameworks tailored to work only at the thread level. I've only seen home-brewed solutions. Please comment if you know of any existing ones.
EDIT:
One could use the lock-free queues from Intel's TBB, either as-is, or as the basis for a more general message-passing queue.
Since you are a beginner, start simple. First make it work correctly, then worry about optimizations. I've seen people try to optimize by increasing the concurrency of a particular section of code (often using dubious tricks), without ever looking to see if there was any contention in the first place.
Second, you want to be able to work at as high a level as you can. Don't work at the level of locks and mutexs if you can using an existing master-worker queue. Intel's TBB looks promising, being slightly higher level than pure threads.
Third, multi-threaded programming is hard. Reduce the areas of your code where you have to think about it as much as possible. If you can write a class such that objects of that class are only ever operated on in a single thread, and there is no static data, it greatly reduces the things that you have to worry about in the class.
A few of the answers have touched on this, but I wanted to emphasize one point:
If you can, make sure that as much of your data as possible is only accessible from one thread at a time. Message queues are a very useful construct to use for this.
I haven't had to write much heavily-threaded code in C++, but in general, the producer-consumer pattern can be very helpful in utilizing multiple threads efficiently, while avoiding the race conditions associated with concurrent access.
If you can use someone else's already-debugged code to handle thread interaction, you're in good shape. As a beginner, there is a temptation to do things in an ad-hoc fashion - to use a "volatile" variable to synchronize between two pieces of code, for example. Avoid that as much as possible. It's very difficult to write code that's bulletproof in the presence of contending threads, so find some code you can trust, and minimize your use of the low-level primitives as much as you can.
My top tips for threading newbies:
If you possibly can, use a task-based parallelism library, Intel's TBB being the most obvious one. This insulates you from the grungy, tricky details and is more efficient than anything you'll cobble together yourself. The main downside is this model doesn't support all uses of multithreading; it's great for exploiting multicores for compute power, less good if you wanted threads for waiting on blocking I/O.
Know how to abort threads (or in the case of TBB, how to make tasks complete early when you decide you didn't want the results after all). Newbies seem to be drawn to thread kill functions like moths to a flame. Don't do it... Herb Sutter has a great short article on this.
Make sure to explicitly know what objects are shared and how they are shared.
As much as possible make your functions purely functional. That is they have inputs and outputs and no side effects. This makes it much simpler to reason about your code. With a simpler program it isn't such a big deal but as the complexity rises it will become essential. Side effects are what lead to thread-safety issues.
Plays devil's advocate with your code. Look at some code and think how could I break this with some well timed thread interleaving. At some point this case will happen.
First learn thread-safety. Once you get that nailed down then you move onto the hard part: Concurrent performance. This is where moving away from global locks is essential. Figuring out ways to minimize and remove locks while still maintaining the thread-safety is hard.
Keep things dead simple as much as possible. It's better to have a simpler design (maintenance, less bugs) than a more complex solution that might have slightly better CPU utilization.
Avoid sharing state between threads as much as possible, this reduces the number of places that must use synchronization.
Avoid false-sharing at all costs (google this term).
Use a thread pool so you're not frequently creating/destroying threads (that's expensive and slow).
Consider using OpenMP, Intel and Microsoft (possibly others) support this extension to C++.
If you are doing number crunching, consider using Intel IPP, which internally uses optimized SIMD functions (this isn't really multi-threading, but is parallelism of a related sorts).
Have tons of fun.
Stay away from MFC and it's multithreading + messaging library.
In fact if you see MFC and threads coming toward you - run for the hills (*)
(*) Unless of course if MFC is coming FROM the hills - in which case run AWAY from the hills.
The biggest "mindset" difference between single-threaded and multi-threaded programming in my opinion is in testing/verification. In single-threaded programming, people will often bash out some half-thought-out code, run it, and if it seems to work, they'll call it good, and often get away with it using it in a production environment.
In multithreaded programming, on the other hand, the program's behavior is non-deterministic, because the exact combination of timing of which threads are running for which periods of time (relative to each other) will be different every time the program runs. So just running a multithreaded program a few times (or even a few million times) and saying "it didn't crash for me, ship it!" is entirely inadequate.
Instead, when doing a multithreaded program, you always should be trying to prove (at least to your own satisfaction) that not only does the program work, but that there is no way it could possibly not work. This is much harder, because instead of verifying a single code-path, you are effectively trying to verify a near-infinite number of possible code-paths.
The only realistic way to do that without having your brain explode is to keep things as bone-headedly simple as you can possibly make them. If you can avoid using multithreading totally, do that. If you must do multithreading, share as little data between threads as possible, and use proper multithreading primitives (e.g. mutexes, thread-safe message queues, wait conditions) and don't try to get away with half-measures (e.g. trying to synchronize access to a shared piece of data using only boolean flags will never work reliably, so don't try it)
What you want to avoid is the multithreading hell scenario: the multithreaded program that runs happily for weeks on end on your test machine, but crashes randomly, about once a year, at the customer's site. That kind of race-condition bug can be nearly impossible to reproduce, and the only way to avoid it is to design your code extremely carefully to guarantee it can't happen.
Threads are strong juju. Use them sparingly.
You should have an understanding of basic systems programing, in particular:
Synchronous vs Asynchronous I/O (blocking vs. non-blocking)
Synchronization mechanisms, such as lock and mutex constructs
Thread management on your target platform
I found viewing the introductory lectures on OS and systems programming here by John Kubiatowicz at Berkeley useful.
Part of my graduate study area relates to parallelism.
I read this book and found it a good summary of approaches at the design level.
At the basic technical level, you have 2 basic options: threads or message passing. Threaded applications are the easiest to get off the ground, since pthreads, windows threads or boost threads are ready to go. However, it brings with it the complexity of shared memory.
Message-passing usability seems mostly limited at this point to the MPI API. It sets up an environment where you can run jobs and partition your program between processors. It's more for supercomputer/cluster environments where there's no intrinsic shared memory. You can achieve similar results with sockets and so forth.
At another level, you can use language type pragmas: the popular one today is OpenMP. I've not used it, but it appears to build threads in via preprocessing or a link-time library.
The classic problem is synchronization here; all the problems in multiprogramming come from the non-deterministic nature of multiprograms, which can not be avoided.
See the Lamport timing methods for a further discussion of synchronizations and timing.
Multithreading is not something that only Ph.D.`s and gurus can do, but you will have to be pretty decent to do it without making insane bugs.
I'm in the same boat as you, I am just starting multi threading for the first time as part of a project and I've been looking around the net for resources. I found this blog to be very informative. Part 1 is pthreads, but I linked starting on the boost section.
I have written a multithreaded server application and a multithreaded shellsort. They were both written in C and use NT's threading functions "raw" that is without any function library in-between to muddle things. They were two quite different experiences with different conclusions to be drawn. High performance and high reliability were the main priorities although coding practices had a higher priority if one of the first two was judged to be threatened in the long term.
The server application had both a server and a client part and used iocps to manage requests and responses. When using iocps it is important never to use more threads than you have cores. Also I found that requests to the server part needed a higher priority so as not to lose any requests unnecessarily. Once they were "safe" I could use lower priority threads to create the server responses. I judged that the client part could have an even lower priority. I asked the questions "what data can't I lose?" and "what data can I allow to fail because I can always retry?" I also needed to be able to interface to the application's settings through a window and it had to be responsive. The trick was that the UI had normal priority, the incoming requests one less and so on. My reasoning behind this was that since I will use the UI so seldom it can have the highest priority so that when I use it it will respond immediately. Threading here turned out to mean that all separate parts of the program in the normal case would/could be running simultaneously but when the system was under higher load, processing power would be shifted to the vital parts due to the prioritization scheme.
I've always liked shellsort so please spare me from pointers about quicksort this or that or blablabla. Or about how shellsort is ill-suited for multithreading. Having said that, the problem I had had to do with sorting a semi-largelist of units in memory (for my tests I used a reverse-sorted list of one million units of forty bytes each. Using a single-threaded shellsort I could sort them at a rate of roughly one unit every two us (microseconds). My first attempt to multithread was with two threads (though I soon realized that I wanted to be able to specify the number of threads) and it ran at about one unit every 3.5 seconds, that is to say SLOWER. Using a profiler helped a lot and one bottleneck turned out to be the statistics logging (i e compares and swaps) where the threads would bump into each other. Dividing up the data between the threads in an efficient way turned out to be the biggest challenge and there is definitley more I can do there such as dividing the vector containing the indeces to the units in cache-line size adapted chunks and perhaps also comparing all indeces in two cache lines before moving to the next line (at least I think there is something I can do there - the algorithms get pretty complicated). In the end, I achieved a rate of one unit every microsecond with three simultaneous threads (four threads about the same, I only had four cores available).
As to the original question my advice to you would be
If you have the time, learn the threading mechanism at the lowest possible level.
If performance is important learn the related mechanisms that the OS provides. Multi-threading by itself is seldom enough to achieve an application's full potential.
Use profiling to understand the quirks of multiple threads working on the same memory.
Sloppy architectural work will kill any app, regardless of how many cores and systems you have executing it and regardless of the brilliance of your programmers.
Sloppy programming will kill any app, regardless of the brilliance of the architectural foundation.
Understand that using libraries lets you reach the development goal faster but at the price of less understanding and (usually) lower performance .
Before giving any advice on do's and dont's about multi-thread programming in C++, I would like to ask the question Is there any particular reason you want to start writing the application in C++?
There are other programming paradigms where you utilize the multi-cores without getting into multi-threaded programming. One such paradigm is functional programming. Write each piece of your code as functions without any side effects. Then it is easy to run it in multiple thread without worrying about synchronization.
I am using Erlang for my development purpose. It has increased by productivity by at least 50%. Code running may not be as fast as the code written in C++. But I have noticed that for most of the back-end offline data processing, speed is not as important as distribution of work and utilizing the hardware as much as possible. Erlang provides a simple concurrency model where you can execute a single function in multiple-threads without worrying about the synchronization issue. Writing multi-threaded code is easy, but debugging that is time consuming. I have done multi-threaded programming in C++, but I am currently happy with Erlang concurrency model. It is worth looking into.
Make sure you know what volatile means and it's uses(which may not be obvious at first).
Also, when designing multithreaded code, it helps to imagine that an infinite amount of processors is executing every single line of code in your application at once. (er, every single line of code that is possible according to your logic in your code.) And that everything that isn't marked volatile the compiler does a special optimization on it so that only the thread that changed it can read/set it's true value and all the other threads get garbage.