Related
As I understand it, a memory barrier will "separate" loads/stores (depending on what type of barrier is used) regardless of the memory address associated with the "fenced" instruction. So if we had an atomic increment, surrounded by loads and stores:
LOAD A
STORE B
LOAD C
LOCK ADD D ; Assume full fence here
LOAD E
STORE F
the instructions operating on A, B and C would have to complete before D; and E and F may not start until after D.
However, as the LOCK is only applied to address D, why restrict the other instructions? Is it too complicated to implement in circuitry? Or is there another reason?
The basic reason is because the basic intent of a fence is to enforce ordering, so if the fence affected only reads/writes of the specific item to which it was applied, it wouldn't do its job.
For example, you fairly frequently have patterns like:
prepare some data
signal that the data is ready
and:
consume some data
signal that the memory used for the data is now free
In such cases, the memory location used as the "signal" is what you're probably going to protect with the fence--but it's not the only thing that really needs to be protected.
In the first case, I have to assure that all the code that writes the data gets executed, and only after it's all done, the signal will be set.
Another thread can then see that the signal is set. Based on that, it knows that it can read all the data associated with the signal, not just the signal itself. If the fence affected only the signal itself, it would mean that the other code that was writing the data might still execute after the signal--and then we'd get a collision between that code writing the data, and the other code trying to read the data.
In theory, we could get around that by using a fence around each individual piece of data being written. In reality, we almost certainly want to avoid that--a fence is fairly expensive, so we'd usually prefer to write a significant amount of data, then use a single fence to indicate that the entire "chunk" of memory is ready.
I am trying to understand the disruptor pattern. I have watched the InfoQ video and tried to read their paper. I understand there is a ring buffer involved, that it is initialized as an extremely large array to take advantage of cache locality, eliminate allocation of new memory.
It sounds like there are one or more atomic integers which keep track of positions. Each 'event' seems to get a unique id and it's position in the ring is found by finding its modulus with respect to the size of the ring, etc., etc.
Unfortunately, I don't have an intuitive sense of how it works. I have done many trading applications and studied the actor model, looked at SEDA, etc.
In their presentation they mentioned that this pattern is basically how routers work; however I haven't found any good descriptions of how routers work either.
Are there some good pointers to a better explanation?
The Google Code project does reference a technical paper on the implementation of the ring buffer, however it is a bit dry, academic and tough going for someone wanting to learn how it works. However there are some blog posts that have started to explain the internals in a more readable way. There is an explanation of ring buffer that is the core of the disruptor pattern, a description of the consumer barriers (the part related to reading from the disruptor) and some information on handling multiple producers available.
The simplest description of the Disruptor is: It is a way of sending messages between threads in the most efficient manner possible. It can be used as an alternative to a queue, but it also shares a number of features with SEDA and Actors.
Compared to Queues:
The Disruptor provides the ability to pass a message onto another threads, waking it up if required (similar to a BlockingQueue). However, there are 3 distinct differences.
The user of the Disruptor defines how messages are stored by extending Entry class and providing a factory to do the preallocation. This allows for either memory reuse (copying) or the Entry could contain a reference to another object.
Putting messages into the Disruptor is a 2-phase process, first a slot is claimed in the ring buffer, which provides the user with the Entry that can be filled with the appropriate data. Then the entry must be committed, this 2-phase approach is necessary to allow for the flexible use of memory mentioned above. It is the commit that makes the message visible to the consumer threads.
It is the responsibility of the consumer to keep track of the messages that have been consumed from the ring buffer. Moving this responsibility away from the ring buffer itself helped reduce the amount of write contention as each thread maintains its own counter.
Compared to Actors
The Actor model is closer the Disruptor than most other programming models, especially if you use the BatchConsumer/BatchHandler classes that are provided. These classes hide all of the complexities of maintaining the consumed sequence numbers and provide a set of simple callbacks when important events occur. However, there are a couple of subtle differences.
The Disruptor uses a 1 thread - 1 consumer model, where Actors use an N:M model i.e. you can have as many actors as you like and they will be distributed across a fixed numbers of threads (generally 1 per core).
The BatchHandler interface provides an additional (and very important) callback onEndOfBatch(). This allows for slow consumers, e.g. those doing I/O to batch events together to improve throughput. It is possible to do batching in other Actor frameworks, however as nearly all other frameworks don't provide a callback at the end of the batch you need to use a timeout to determine the end of the batch, resulting in poor latency.
Compared to SEDA
LMAX built the Disruptor pattern to replace a SEDA based approach.
The main improvement that it provided over SEDA was the ability to do work in parallel. To do this the Disruptor supports multi-casting the same messages (in the same order) to multiple consumers. This avoids the need for fork stages in the pipeline.
We also allow consumers to wait on the results of other consumers without having to put another queuing stage between them. A consumer can simply watch the sequence number of a consumer that it is dependent on. This avoids the need for join stages in pipeline.
Compared to Memory Barriers
Another way to think about it is as a structured, ordered memory barrier. Where the producer barrier forms the write barrier and the consumer barrier is the read barrier.
First we'd like to understand the programming model it offers.
There are one or more writers. There are one or more readers. There is a line of entries, totally ordered from old to new (pictured as left to right). Writers can add new entries on the right end. Every reader reads entries sequentially from left to right. Readers can't read past writers, obviously.
There is no concept of entry deletion. I use "reader" instead of "consumer" to avoid the image of entries being consumed. However we understand that entries on the left of the last reader become useless.
Generally readers can read concurrently and independently. However we can declare dependencies among readers. Reader dependencies can be arbitrary acyclic graph. If reader B depends on reader A, reader B can't read past reader A.
Reader dependency arises because reader A can annotate an entry, and reader B depends on that annotation. For example, A does some calculation on an entry, and stores the result in field a in the entry. A then move on, and now B can read the entry, and the value of a A stored. If reader C does not depend on A, C should not attempt to read a.
This is indeed an interesting programming model. Regardless of the performance, the model alone can benefit lots of applications.
Of course, LMAX's main goal is performance. It uses a pre-allocated ring of entries. The ring is large enough, but it's bounded so that the system will not be loaded beyond design capacity. If the ring is full, writer(s) will wait until the slowest readers advance and make room.
Entry objects are pre-allocated and live forever, to reduce garbage collection cost. We don't insert new entry objects or delete old entry objects, instead, a writer asks for a pre-existing entry, populate its fields, and notify readers. This apparent 2-phase action is really simply an atomic action
setNewEntry(EntryPopulator);
interface EntryPopulator{ void populate(Entry existingEntry); }
Pre-allocating entries also means adjacent entries (very likely) locate in adjacent memory cells, and because readers read entries sequentially, this is important to utilize CPU caches.
And lots of efforts to avoid lock, CAS, even memory barrier (e.g. use a non-volatile sequence variable if there's only one writer)
For developers of readers: Different annotating readers should write to different fields, to avoid write contention. (Actually they should write to different cache lines.) An annotating reader should not touch anything that other non-dependent readers may read. This is why I say these readers annotate entries, instead of modify entries.
Martin Fowler has written an article about LMAX and the disruptor pattern, The LMAX Architecture, which may clarify it further.
I actually took the time to study the actual source, out of sheer curiosity, and the idea behind it is quite simple. The most recent version at the time of writing this post is 3.2.1.
There is a buffer storing pre-allocated events that will hold the data for consumers to read.
The buffer is backed by an array of flags (integer array) of its length that describes the availability of the buffer slots (see further for details). The array is accessed like a java#AtomicIntegerArray, so for the purpose of this explenation you may as well assume it to be one.
There can be any number of producers. When the producer wants to write to the buffer, an long number is generated (as in calling AtomicLong#getAndIncrement, the Disruptor actually uses its own implementation, but it works in the same manner). Let's call this generated long a producerCallId. In a similar manner, a consumerCallId is generated when a consumer ENDS reading a slot from a buffer. The most recent consumerCallId is accessed.
(If there are many consumers, the call with the lowest id is choosen.)
These ids are then compared, and if the difference between the two is lesser that the buffer side, the producer is allowed to write.
(If the producerCallId is greater than the recent consumerCallId + bufferSize, it means that the buffer is full, and the producer is forced to bus-wait until a spot becomes available.)
The producer is then assigned the slot in the buffer based on his callId (which is prducerCallId modulo bufferSize, but since the bufferSize is always a power of 2 (limit enforced on buffer creation), the actuall operation used is producerCallId & (bufferSize - 1)). It is then free to modify the event in that slot.
(The actual algorithm is a bit more complicated, involving caching recent consumerId in a separate atomic reference, for optimisation purposes.)
When the event was modified, the change is "published". When publishing the respective slot in the flag array is filled with the updated flag. The flag value is the number of the loop (producerCallId divided by bufferSize (again since bufferSize is power of 2, the actual operation is a right shift).
In a similar manner there can be any number of consumers. Each time a consumer wants to access the buffer, a consumerCallId is generated (depending on how the consumers were added to the disruptor the atomic used in id generation may be shared or separate for each of them). This consumerCallId is then compared to the most recent producentCallId, and if it is lesser of the two, the reader is allowed to progress.
(Similarly if the producerCallId is even to the consumerCallId, it means that the buffer is empety and the consumer is forced to wait. The manner of waiting is defined by a WaitStrategy during disruptor creation.)
For individual consumers (the ones with their own id generator), the next thing checked is the ability to batch consume. The slots in the buffer are examined in order from the one respective to the consumerCallId (the index is determined in the same manner as for producers), to the one respective to the recent producerCallId.
They are examined in a loop by comparing the flag value written in the flag array, against a flag value generated for the consumerCallId. If the flags match it means that the producers filling the slots has commited their changes. If not, the loop is broken, and the highest commited changeId is returned. The slots from ConsumerCallId to received in changeId can be consumed in batch.
If a group of consumers read together (the ones with shared id generator), each one only takes a single callId, and only the slot for that single callId is checked and returned.
From this article:
The disruptor pattern is a batching queue backed up by a circular
array (i.e. the ring buffer) filled with pre-allocated transfer
objects which uses memory-barriers to synchronize producers and
consumers through sequences.
Memory-barriers are kind of hard to explain and Trisha's blog has done the best attempt in my opinion with this post: http://mechanitis.blogspot.com/2011/08/dissecting-disruptor-why-its-so-fast.html
But if you don't want to dive into the low-level details you can just know that memory-barriers in Java are implemented through the volatile keyword or through the java.util.concurrent.AtomicLong. The disruptor pattern sequences are AtomicLongs and are communicated back and forth among producers and consumers through memory-barriers instead of locks.
I find it easier to understand a concept through code, so the code below is a simple helloworld from CoralQueue, which is a disruptor pattern implementation done by CoralBlocks with which I am affiliated. In the code below you can see how the disruptor pattern implements batching and how the ring-buffer (i.e. circular array) allows for garbage-free communication between two threads:
package com.coralblocks.coralqueue.sample.queue;
import com.coralblocks.coralqueue.AtomicQueue;
import com.coralblocks.coralqueue.Queue;
import com.coralblocks.coralqueue.util.MutableLong;
public class Sample {
public static void main(String[] args) throws InterruptedException {
final Queue<MutableLong> queue = new AtomicQueue<MutableLong>(1024, MutableLong.class);
Thread consumer = new Thread() {
#Override
public void run() {
boolean running = true;
while(running) {
long avail;
while((avail = queue.availableToPoll()) == 0); // busy spin
for(int i = 0; i < avail; i++) {
MutableLong ml = queue.poll();
if (ml.get() == -1) {
running = false;
} else {
System.out.println(ml.get());
}
}
queue.donePolling();
}
}
};
consumer.start();
MutableLong ml;
for(int i = 0; i < 10; i++) {
while((ml = queue.nextToDispatch()) == null); // busy spin
ml.set(System.nanoTime());
queue.flush();
}
// send a message to stop consumer...
while((ml = queue.nextToDispatch()) == null); // busy spin
ml.set(-1);
queue.flush();
consumer.join(); // wait for the consumer thread to die...
}
}
I am looking for a method to implement lock-free queue data structure that supports single producer, and multiple consumers. I have looked at the classic method by Maged Michael and Michael Scott (1996) but their version uses linked lists. I would like an implementation that makes use of bounded circular buffer. Something that uses atomic variables?
On a side note, I am not sure why these classic methods are designed for linked lists that require a lot of dynamic memory management. In a multi-threaded program, all memory management routines are serialized. Aren't we defeating the benefits of lock-free methods by using them in conjunction with dynamic data structures?
I am trying to code this in C/C++ using pthread library on a Intel 64-bit architecture.
Thank you,
Shirish
The use of a circular buffer makes a lock necessary, since blocking is needed to prevent the head from going past the tail. But otherwise the head and tail pointers can easily be updated atomically. Or in some cases the buffer can be so large that overwriting is not an issue. (in real life you will see this in automated trading systems, with circular buffers sized to hold X minutes of market data. If you are X minutes behind, you have wayyyy worse problems than overwriting your buffer).
When I implemented the MS queue in C++, I built a lock free allocator using a stack, which is very easy to implement. If I have MSQueue then at compile time I know sizeof(MSQueue::node). Then I make a stack of N buffers of the required size. The N can grow, i.e. if pop() returns null, it is easy to go ask the heap for more blocks, and these are pushed onto the stack. Outside of the possibly blocking call for more memory, this is a lock free operation.
Note that the T cannot have a non-trivial dtor. I worked on a version that did allow for non-trivial dtors, that actually worked. But I found that it was easier just to make the T a pointer to the T that I wanted, where the producer released ownership, and the consumer acquired ownership. This of course requires that the T itself is allocated using lockfree methods, but the same allocator I made with the stack works here as well.
In any case the point of lock-free programming is not that the data structures themselves are slower. The points are this:
lock free makes me independent of the scheduler. Lock-based programming depends on the scheduler to make sure that the holders of a lock are running so that they can release the lock. This is what causes "priority inversion" On Linux there are some lock attributes to make sure this happens
If I am independent of the scheduler, the OS has a far easier time managing timeslices, and I get far less context switching
it is easier to write correct multithreaded programs using lockfree methods since I dont have to worry about deadlock , livelock, scheduling, syncronization, etc This is espcially true with shared memory implementations, where a process could die while holding a lock in shared memory, and there is no way to release the lock
lock free methods are far easier to scale. In fact, I have implemented lock free methods using messaging over a network. Distributed locks like this are a nightmare
That said, there are many cases where lock-based methods are preferable and/or required
when updating things that are expensive or impossible to copy. Most lock free methods use some sort of versioning, i.e. make a copy of the object, update it, and check if the shared version is still the same as when you copied it, then make the current version you update version. Els ecopy it again, apply the update, and check again. Keep doing this until it works. This is fine when the objects are small, but it they are large, or contain file handles, etc then not recommended
Most types are impossible to access in a lock free way, e.g. any STL container. These have invariants that require non atomic access , for example assert(vector.size()==vector.end()-vector.begin()). So if you are updating/reading a vector that is shared, you have to lock it.
This is an old question, but no one has provided an accepted solution. So I offer this info for others who may be searching.
This website: http://www.1024cores.net
Provides some really useful lockfree/waitfree data structures with thorough explanations.
What you are seeking is a lock-free solution to the reader/writer problem.
See: http://www.1024cores.net/home/lock-free-algorithms/reader-writer-problem
For a traditional one-block circular buffer I think this simply cannot be done safely with atomic operations. You need to do so much in one read. Suppose you have a structure that has this:
uint8_t* buf;
unsigned int size; // Actual max. buffer size
unsigned int length; // Actual stored data length (suppose in write prohibited from being > size)
unsigned int offset; // Start of current stored data
On a read you need to do the following (this is how I implemented it anyway, you can swap some steps like I'll discuss afterwards):
Check if the read length does not surpass stored length
Check if the offset+read length do not surpass buffer boundaries
Read data out
Increase offset, decrease length
What should you certainly do synchronised (so atomic) to make this work? Actually combine steps 1 and 4 in one atomic step, or to clarify: do this synchronised:
check read_length, this can be sth like read_length=min(read_length,length);
decrease length with read_length: length-=read_length
get a local copy from offset unsigned int local_offset = offset
increase offset with read_length: offset+=read_length
Afterwards you can just do a memcpy (or whatever) starting from your local_offset, check if your read goes over circular buffer size (split in 2 memcpy's), ... . This is 'quite' threadsafe, your write method could still write over the memory you're reading, so make sure your buffer is really large enough to minimize that possibility.
Now, while I can imagine you can combine 3 and 4 (I guess that's what they do in the linked-list case) or even 1 and 2 in atomic operations, I cannot see you do this whole deal in one atomic operation :).
You can however try to drop 'length' checking if your consumers are very smart and will always know what to read. You'd also need a new woffset variable then, because the old method of (offset+length)%size to determine write offset wouldn't work anymore. Note this is close to the case of a linked list, where you actually always read one element (= fixed, known size) from the list. Also here, if you make it a circular linked list, you can read to much or write to a position you're reading at that moment!
Finally: my advise, just go with locks, I use a CircularBuffer class, completely safe for reading & writing) for a realtime 720p60 video streamer and I have got no speed issues at all from locking.
This is an old question but no one has provided an answer that precisely answers it. Given that still comes up high in search results for (nearly) the same question, there should be an answer, given that one exists.
There may be more than one solution, but here is one that has an implementation:
https://github.com/tudinfse/FFQ
The conference paper referenced in the readme details the algorithm.
On Linux I'm using shmget and shmat to setup a shared memory segment that one process will write to and one or more processes will read from. The data that is being shared is a few megabytes in size and when updated is completely rewritten; it's never partially updated.
I have my shared memory segment laid out as follows:
-------------------------
| t0 | actual data | t1 |
-------------------------
where t0 and t1 are copies of the time when the writer began its update (with enough precision such that successive updates are guaranteed to have differing times). The writer first writes to t1, then copies in the data, then writes to t0. The reader on the other hand reads t0, then the data, then t1. If the reader gets the same value for t0 and t1 then it considers the data consistent and valid, if not, it tries again.
Does this procedure ensure that if the reader thinks the data is valid then it actually is?
Do I need to worry about out-of-order execution (OOE)? If so, would the reader using memcpy to get the entire shared memory segment overcome the OOE issues on the reader side? (This assumes that memcpy performs it's copy linearly and ascending through the address space. Is that assumption valid?)
Modern hardware is actually anything but sequentially consistent. Thus, this is not guaranteed to work as such if you don't execute memory barriers at the appropriate spots. Barriers are needed because the architecture implements a weaker shared memory consistency model than sequential consistency. This as such has nothing to do with pipelining or OoO, but with allowing multiple processors to efficiently access the memory system in parallel. See e.g. Shared memory consistency models: A tutorial. On a uniprocessor, you don't need barriers, because all the code executes sequentially on that one processor.
Also, there is no need to have two time fields, a sequence count is probably a better choice as there is no need to worry whether two updates are so close that they get the same timestamp, and updating a counter is much faster than getting the current time. Also, there is no chance that the clock moves backwards in time which might happen e.g. when ntpd adjusts for clock drift. Though this last problem can be overcome on Linux by using clock_gettime(CLOCK_MONOTONIC, ...). Another advantage of using sequence counters instead of timestamps is that you need only one sequence counter. The writer increments the counter both before writing the data, and after the write is done. Then the reader reads the sequence number, checks that it's even, and if so, reads the data, and finally then reads the sequence number again and compares to the first sequence number. If the sequence number is odd, it means a write is in progress, and there is no need to read the data.
The Linux kernel uses a locking primitive called seqlocks that do something like the above. If you're not afraid of "GPL contamination", you can google for the implementation; As such it's trivial, but the trick is getting the barriers correct.
Joe Duffy gives the exact same algorithm and calls it: "A scalable reader/writer scheme with optimistic retry".
It works.
You need two sequence number fields.
You need to read and write them in opposite order.
You might need to have memory barriers in place, depending on the memory ordering guarantees of the system.
Specifically, you need read acquire and store release semantics for the readers and writers when they access t0 or t1 for reading and writing respectively.
What instructions are needed to achieve this, depends on the architecture. E.g. on x86/x64, because of the relatively strong guarantees one needs no machine specific barriers at all in this specific case*.
* one still needs to ensure that the compiler/JIT does not mess around with loads and stores , e.g. by using volatile (that has a different meaning in Java and C# than in ISO C/C++. Compilers may differ, however. E.g. using VC++ 2005 or above with volatile it would be safe doing the above. See the "Microsoft Specific" section. It can be done with other compilers as well on x86/x64. The assembly code emitted should be inspected and one must make sure that accesses to t0 and t1 are not eliminated or moved around by the compiler.)
As a side note, if you ever need MFENCE, lock or [TopOfStack],0 might be a better option, depending on your needs.
I have a large array of structures, like this:
typedef struct
{
int a;
int b;
int c;
etc...
}
data_type;
data_type data[100000];
I have a bunch of separate threads, each of which will want to make alterations to elements within data[]. I need to make sure that no to threads attempt to access the same data element at the same time. To be precise: one thread performing data[475].a = 3; and another thread performing data[475].b = 7; at the same time is not allowed, but one thread performing data[475].a = 3; while another thread performs data[476].a = 7; is allowed. The program is highly speed critical. My plan is to make a separate critical section for each data element like so:
typedef struct
{
CRITICAL_SECTION critsec;
int a;
int b;
int c;
etc...
}
data_type;
In one way I guess it should all work and I should have no real questions, but not having had much experience in multithreaded programming I am just feeling a little uneasy about having so many critical sections. I'm wondering if the sheer number of them could be creating some sort of inefficiency. I'm also wondering if perhaps some other multithreading technique could be faster? Should I just relax and go ahead with plan A?
With this many objects, most of their critical sections will be unlocked, and there will be almost no contention. As you already know (other comment), critical sections don't require a kernel-mode transition if they're unowned. That makes critical sections efficient for this situation.
The only other consideration would be whether you would want the critical sections inside your objects or in another array. Locality of reference is a good reason to put the critical sections inside the object. When you've entered the critical section, an entire cacheline (e.g. 16 or 32 bytes) will be in memory. With a bit of padding, you can make sure each object starts on a cacheline. As a result, the object will be (partially) in cache once its critical section is entered.
Your plan is worth trying, but I think you will find that Windows is unhappy creating that many Critical Sections. Each CS contains some kernel handle(s) and you are using up precious kernel space. I think, depending on your version of Windows, you will run out of handle memory and InitializeCriticalSection() or some other function will start to fail.
What you might want to do is have a pool of CSs available for use, and store a pointer to the 'in use' CS inside your struct. But then this gets tricky quite quickly and you will need to use Atomic operations to set/clear the CS pointer (to atomically flag the array entry as 'in use'). Might also need some reference counting, etc...
Gets complicated.
So try your way first, and see what happens. We had a similar situation once, and we had to go with a pool, but maybe things have changed since then.
Depending on the data member types in your data_type structure (and also depending on the operations you want to perform on those members), you might be able to forgo using a separate synchronization object, using the Interlocked functions instead.
In your sample code, all the data members are integers, and all the operations are assignments (and presumably reads), so you could use InterlockedExchange() to set the values atomically and InterlockedCompareExchange() to read the values atomically.
If you need to use non-integer data member types, or if you need to perform more complex operations, or if you need to coordinate atomic access to more than one operation at a time (e.g., read data[1].a and then write data[1].b), then you will have to use a synchronization object, such as a CRITICAL_SECTION.
If you must use a synchronization object, I recommend that you consider partitioning your data set into subsets and use a single synchronization object per subset. For example, you might consider using one CRITICAL_SECTION for each span of 1000 elements in the data array.
You could also consider MUTEX.
This is nice method.
Each client could reserve the resource by itself with mutex (mutual-exclusion).
This is more common, some libraries also support this with threads.
Read about boost::thread and it's mutexes
With Your approach:
data_type data[100000];
I'd be afraid of stack overflow, unless You're allocating it at the heap.
EDIT:
Boost::MUTEX
uses win32 Critical Sections
As others have pointed out, yes there is an issue and it is called too fine-grained locking.. it's resource wasteful and even though the chances are small you will start creating a lot of backing primitives and data when the things do get an occasional, call it longer than usual or whatever, contention. Plus you are wasting resources as it is not really a trivial data structure as for example in VM impls..
If I recall correctly you will have a higher chance of a SEH exception from that point onwards on Win32 or just higher memory usage. Partitioning and pooling them is probably the way to go but it is a more complex implementation. Paritioning on something else (re:action) and expecting some short-lived contention is another way to deal with it.
In any case, it is a problem of resource management with what you have right now.