Initializing Billion Integers to value 1 - c++

What is good posix thread design to initialize billion integers using c/c++ on linux platform 8-core CPU with 32GB of DRAM?
Thanks for your help.

This is a trivial operation and you need not consider multi-threading. Just do it with a memcpy in a single thread.

The exact number of threads will not be such a limiting factor, but sometimes for this questions it is worth to overcommit, say use 2 threads per physical core.
But the real bottleneck will be IO, writing the data into the RAM. You'd have to take care that the data that is to be replaced will never read before you erase it. Then you should assure that writes to memory appear in large chunks and (if possible) as "write through", mondern CPU have instructions for the later.
Usually something like memcpy with a fixed sized buffer (some pages) that contains the pattern that you want to see should be optimized quite well.

What is that for? Depending on usage, the following scenario might work: you initialize one memory page (that's several KB) to all 1's. Then you map that page into the virtual address space as many times as needed with a copy-on-write flag. This way, on reading you'll get all ones from all those virtual pages, on writing the system will allocate more physical pages as needed.

Perhaps a divide and conquer algorithm? Partition the memory containing the integers by some number corresponding to the number of threads optimal for your system. Then launch one thread per partition which initializes all of its integers.

If you do attempt multithreading, aligning your writes with the native cache line size will likely provide optimal memory throughput. As everyone says, the memory throughput will dominate the performance but there is some portion of CPU time required for these writes. Minimizing that time with multithreading and vectorized instructions may be helpful.
The real answer is to profile your system (since you stated a very specific target, it sounds like you don't want to design a balanced algorithm which is good enough for most targets). Modern CPUs which have access to 32GB of DRAM often have hardware performance counters (Intel's and AMD's do) which make finding out CPU, caching activity pretty easy.

Related

Hard disk contention using multiple threads

I have not performed any profile testing of this yet, but what would the general consensus be on the advantages/disadvantages of resource loading from the hard disk using multiple threads vs one thread? Note. I am not talking about the main thread.
I would have thought that using more than one "other" thread to do the loading to be pointless because the HD cannot do 2 things at once, and therefore would surely only cause disk contention.
Not sure which way to go architecturally, appreciate any advice.
EDIT: Apologies, I meant to mean an SSD drive not a magnetic drive. Both are HD's to me, but I am more interested in the case of a system with a single SSD drive.
As pointed out in the comments one advantage of using multiple threads is that a large file load will not delay the presentation of a smaller for to the receiver of the thread loader. In my case, this is a big advantage, and so even if it costs a little perf to do it, having multiple threads is desirable.
I know there are no simple answers, but the real question I am asking is, what kind of performance % penalty would there be for making the parallel disk writes sequential (in the OS layer) as opposed to allowing only 1 resource loader thread? And what are the factors that drive this? I don't mean like platform, manufacturer etc. I mean technically, what aspects of the OS/HD interaction influence this penalty? (in theory).
FURTHER EDIT:
My exact use case are texture loading threads which only exist to load from HD and then "pass" them on to opengl, so there is minimal "computation in the threads (maybe some type conversion etc). In this case, the thread would spend most of its time waiting for the HD (I would of thought), and therefore how the OS-HD interaction is managed is important to understand. My OS is Windows 10.
Note. I am not talking about the main thread.
Main vs non-main thread makes zero difference to the speed of reading a disk.
I would have thought that using more than one "other" thread to do the loading to be pointless because the HD cannot do 2 things at once, and therefore would surely only cause disk contention.
Indeed. Not only are the attempted parallel reads forced to wait for each other (and thus not actually be parallel), but they will also make access pattern of the disk random as opposed to sequential, which is much much slower due to disk head seek time.
Of course, if you were to deal with multiple hard disks, then one thread dedicated for each drive would probably be optimal.
Now, if you were using a solid state drive instead of a hard drive, the situation isn't quite so clear cut. Multiple threads may be faster, slower, or comparable. There are probably many factors involved such as firmware, file system, operating system, speed of the drive relative to some other bottle neck, etc.
In either case, RAID might invalidate assumptions made here.
It depends on how much processing of the data you're going to do. This will determine whether the application is I/O you bound or compute bound.
For example, if all you are going to do to the data is some simple arithmetic, e.g. add 1, then you will end up being I/O bound. The CPU can add 1 to data far quicker than any I/O system can deliver flows of data.
However, if you're going to do a large amount of work on each batch of data, e.g. a FFT, then a filter, then a convolution (I'm picking random DSP routine names here), then it's likely that you will end up being compute bound; the CPU cannot keep up with the data being delivered by the I/O subsystem which owns your SSD.
It is quite an art to judge just how an algorithm should be structured to match the underlying capabilities of the underlying machine, and vice versa. There's profiling tools like FTRACE/Kernelshark, Intel's VTune, which are both useful in analysing exactly what is going on. Google does a lot to measure how many searches-per-Watt their hardware accomplishes, power being their biggest cost.
In general I/O of any sort, even a big array of SSDs, is painfully slow. Even the main memory in a PC (DDR4) is painfully slow in comparison to what the CPU can consume. Even the L3 and L2 caches are sluggards in comparison to the CPU cores. It's hard to design and multi-threadify an algorithm just right so that the right amount of work is done on each data item whilst it is in L1 cache so that the L2, L3 caches, DDR4 and I/O subsystems can deliver the next data item to the L1 caches just in time to keep the CPU cores busy. And the ideal software design for one machine is likely hopeless on another with a different CPU, or SSD, or memory SIMMs. Intel design for good general purpose computer performance, and actually extracting peak performance from a single program is a real challenge. Libraries like Intel's MKL and IPP are very big helps in doing this.
General Guidance
In general one should look at it in terms of data bandwidth required by any particular arrangement of threads and work those threads are doing.
This means benchmarking your program's inner processing loop and measuring how much data it processed and how quickly it managed to do it in, choosing an number of data items that makes sense but much more than the size of L3 cache. A single 'data item' is an amount of input data, the amount of corresponding output data, and any variables used processing the input to the output, the total size of which fits in L1 cache (with some room to spare). And no cheating - use the CPUs SSE/AVX instructions where appropriate, don't forego them by writing plain C or not using something like Intel's IPP/MKL. [Though if one is using IPP/MKL, it kinda does all this for you to the best of its ability.]
These days DDR4 memory is going to be good for anything between 20 to 100GByte/second (depending on what CPU, number of SIMMs, etc), so long as your not making random, scattered accesses to the data. By saturating the L3 your are forcing yourself into being bound by the DDR4 speed. Then you can start changing your code, increasing the work done by each thread on a single data item. Keep increasing the work per item and the speed will eventually start increasing; you've reached the point where you are no longer limited by the speed of DDR4, then L3, then L2.
If after this you can still see ways of increasing the work per data item, then keep going. You eventually get to a data bandwidth somewhere near that of the IO subsystems, and only then will you be getting the absolute most out of the machine.
It's an iterative process, and experience allows one to short cut it.
Of course, if one runs out of ideas for things to increase the work done per data item then that's the end of the design process. More performance can be achieved only by improving the bandwidth of whatever has ended up being the bottleneck (almost certainly the SSD).
For those of us who like doing this software of thing, the PS3's Cell processor was a dream. No need to second guess the cache, there was none. One had complete control over what data and code was where and when it was there.
A lot people will tell you that an HD can't do more than one thing at once. This isn't quite true because modern IO systems have a lot of indirection. Saturating them is difficult to do with one thread.
Here are three scenarios that I have experienced where multi-threading the IO helps.
Sometimes the IO reading library has a non-trivial amount of computation, think about reading compressed videos, or parity checking after the transfer has happened. One example is using robocopy with multiple threads. Its not unusual to launch robocopy with 128 threads!
Many operating systems are designed so that a single process can't saturate the IO, because this would lead to system unresponsiveness. In one case I got a 3% percent read speed improvement because I came closer to saturating the IO. This is doubly true if some system policy exists to stripe the data to different drives, as might be set on a Lustre drive in a HPC cluster. For my application, the optimal number of threads was two.
More complicated IO, like a RAID card, contains a substantial cache that keep the HD head constantly reading and writing. To get optimal throughput you need to be sure that whenever the head is spinning its constantly reading/writing and not just moving. The only way to do this is, in practice, is to saturate the card's on-board RAM.
So, many times you can overlap some minor amount of computation by using multiple threads, and stuff starts getting tricky with larger disk arrays.
Not sure which way to go architecturally, appreciate any advice.
Determining the amount of work per thread is the most common architectural optimization. Write code so that its easy to increase the IO worker count. You're going to need to benchmark.

Will multithreading improve performance significantly if I have a fixed amount of calculations that are independet from each other?

I am programming a raycasting game engine.
Each ray can be calculated without knowing anything about the other rays (I'm only calculating distances).
Since there is no waiting time between calculations, I wonder whether it's worth the effort to make the ray calculations multithreaded or not.
Is it likely that there will be a performance boost?
Mostly likely multi-threading will improve performance if done correctly. The way you've stated your problem, it is a perfect candidate for multi-threading since the computations are independent, reducing the need for coordination between threads to a minimum.
Some reasons you still might not get a speed up, or may not get the full speed up you expect could include:
1) The bottleneck may not be on-die CPU execution resources (e.g., ALU-bound operations), but rather something shared like memory or shared LLC bandwidth.
For example, on some architectures, a single thread may be able to saturate memory bandwidth, so adding more cores may not help. A more common case is that a single core can saturate some fraction, 1/N < 1 of main memory bandwidth, and this value is larger than 1/C where C is the core count. For instance, on a 4 core box, one core may be able to consume 50% of the bandwidth. Then, for a memory-bound computation, you'll get good scaling to 2 cores (using 100% of bandwidth), but little to none above that.
Other resources which are shared among cores include disk and network IO, GPU, snoop bandwidth, etc. If you have a hyper-threaded platform, this list increases to include all levels of cache and ALU resources for logical cores sharing the same physical core.
2) Contention "in practice" between operations which are "theoretically" independent.
You mention that your operations are independent. Typically this means that they are logically independent - they don't share any data (other than perhaps immutable input) and they can write to separate output areas. That doesn't exclude the possibility, however, than any given implementation actually has some hidden sharing going on.
One classic example is false-sharing - where independent variables fall in the same cache line, so logically independent writes to different variables from different threads end up thrashing the cache line between cores.
Another example, frequently encountered in practice, is contention via library - if your routines use malloc heavily, you may find that all the threads spend most of their time waiting on a lock inside the allocator as malloc is shared resource. This can be remedied by reducing reliance on malloc (perhaps via fewer, larger mallocs) or with a good concurrent malloc such as hoard or tcmalloc.
3) Implementation of the distribution and collection of the computation across threads may overwhelm the advantage you get from multiple threads. For example, if you spin up a new thread for every individual ray, the thread creation overhead would dominate your runtime and you would likely see a negative benefit. Even if you use a thread-pool of persistent threads, choosing a "work unit" that is too fine grained will impose a lot of coordination overhead which may eliminate your benefits.
Similarly, if you have to copy the input data to and from the worker threads, you may not see the scaling you expect. Where possible, use pass-by-reference for read-only data.
4) You don't have more than 1 core, or you do have more than 1 core but they are already occupied running other threads or processes. In these cases, the effort to coordinate multiple threads is pure overhead.
In general, it depends. Given that the calculations are independent, it sounds like this is a good candidate for potential performance improvements due to threading. Ray calculations typically can benefit from this.
However, there are many other factors, such as memory access requirements, as well as the underlying system on which this runs, which will have a tremendous impact on this. It's often possible to have multithreaded versions run slower than single threaded versions if not written correctly, so profiling is the only way to answer this definitively.
Probably yes, multithreading (e.g. with pthreads) could improve performance; but you surely want to benchmark (and you might be disappointed if your program is memory bound, not CPU bound). And you could also consider OpenCL (to run some regular numeric computations on the GPGPU) and OpenMP (to explicitly ask the compiler, thru pragmas, to parallelize some of your code).
Maybe Open-MPI might be considered to run on several communicating processes. And if you are brave (or crazy) you could mix several approaches.
In reality, it depends upon the algorithm and the system (both hardware and operating system), and you should benchmark (e.g. some micro-prototype related to your needs).
If on some particular system the bottleneck is the memory bandwidth (not the CPU), multi-threading or multi-processing won't help much (and probably could degrade performance).
Also, the cost of synchronization may vary widely (e.g. locking a mutex can be very fast on some systems and 50x slower on others).
Very likely. Independent calculations are a perfect candidate for parallelization. In the case of raycasting, there is so many of them that they would spread nicely across as many parallel threads as the hardware permits.
An unexpected bottleneck for calculations that would otherwise have perfect data-independence can be concurrent writes to nearby locations (false sharing of cache lines).

Optimal buffer size for write(2)

Let's say I want to write a 1 GB of data to the file on, say ext3 Linux filesystem using write(2) syscall and this happens in a very busy environment (many similar I/Os concurently). What is the optimal buffer size in the interval, say, [4 kB, 4 MB] to do that when
not using O_DIRECT open flag, or
using O_DIRECT?
Please, no "check it yourself" answers -- I'd like to get some answer from "filesystems" guys.
The answer is in my experience much more dependent on the underlying devices and hardware rather than the filesystem itself -- that is buffer caches on the device, and the capabilities of the device to write in small blocks etc -- however you should never write in smaller sizes than your file system block size (stat(.) -- likely to be about 4kb) -- similarly you should not really go beyond the L2/L3 cache size of the CPU which in many cases can be as low as 512kb.
Given that SSD devices and similar like the 64kb as the unit of operations, then I would suggest that a buffer size of 64kb-128kb being the most optimal -- which also correspond with my empirical experience as having the highest throughput.
As discussed in comments, I believe the exact size don't matter that much, assuming it is :
a small multiple of the file system size (see comment by Joachim Pileborg suggesting stat(".") etc.)
a power of two (because computers and kernels like them)
not too big (e.g. fitting in some cache inside your processor, e.g. L2 cache)
aligned in memory (e.g. to a page size using posix_memalign).
So a power of two between 16kbytes and a few megabytes should probably fit. Most of the time is spent on reading the disk. Filesystem and disk benchmarks are quite flat in that range.
4Kbytes seems to often be the page size and the disk chunk size.
Of course, you can tune things, even tune, when making the file system with mke2fs, the file system block size.
And I'll bet that the optimal is really dependent upon your hardware (SSD, hard disks?) and your system (and its load).
Including stdio.h should define BUFSIZ as the optimal size for the system. This is by no means guaranteed, but it is the right value to use if you do not have the ability to do extensive benchmarks, and it is a good starting point for such benchmarks.

C++ cache aware programming

is there a way in C++ to determine the CPU's cache size? i have an algorithm that processes a lot of data and i'd like to break this data down into chunks such that they fit into the cache. Is this possible?
Can you give me any other hints on programming with cache-size in mind (especially in regard to multithreaded/multicore data processing)?
Thanks!
According to "What every programmer should know about memory", by Ulrich Drepper you can do the following on Linux:
Once we have a formula for the memory
requirement we can compare it with the
cache size. As mentioned before, the
cache might be shared with multiple
other cores. Currently {There
definitely will sometime soon be a
better way!} the only way to get
correct information without hardcoding
knowledge is through the /sys
filesystem. In Table 5.2 we have seen
the what the kernel publishes about
the hardware. A program has to find
the directory:
/sys/devices/system/cpu/cpu*/cache
This is listed in Section 6: What Programmers Can Do.
He also describes a short test right under Figure 6.5 which can be used to determine L1D cache size if you can't get it from the OS.
There is one more thing I ran across in his paper: sysconf(_SC_LEVEL2_CACHE_SIZE) is a system call on Linux which is supposed to return the L2 cache size although it doesn't seem to be well documented.
C++ itself doesn't "care" about CPU caches, so there's no support for querying cache-sizes built into the language. If you are developing for Windows, then there's the GetLogicalProcessorInformation()-function, which can be used to query information about the CPU caches.
Preallocate a large array. Then access each element sequentially and record the time for each access. Ideally there will be a jump in access time when cache miss occurs. Then you can calculate your L1 Cache. It might not work but worth trying.
read the cpuid of the cpu (x86) and then determine the cache-size by a look-up-table. The table has to be filled with the cache sizes the manufacturer of the cpu publishes in its programming manuals.
Depending on what you're trying to do, you might also leave it to some library. Since you mention multicore processing, you might want to have a look at Intel Threading Building Blocks.
TBB includes cache aware memory allocators. More specifically, check cache_aligned_allocator (in the reference documentation, I couldn't find any direct link).
Interestingly enough, I wrote a program to do this awhile ago (in C though, but I'm sure it will be easy to incorporate in C++ code).
http://github.com/wowus/CacheLineDetection/blob/master/Cache%20Line%20Detection/cache.c
The get_cache_line function is the interesting one, which returns the location of right before the biggest spike in timing data of array accesses. It correctly guessed on my machine! If anything else, it can help you make your own.
It's based off of this article, which originally piqued my interest: http://igoro.com/archive/gallery-of-processor-cache-effects/
You can see this thread: http://software.intel.com/en-us/forums/topic/296674
The short answer is in this other thread:
On modern IA-32 hardware, the cache line size is 64. The value 128 is
a legacy of the Intel Netburst Microarchitecture (e.g. Intel Pentium
D) where 64-byte lines are paired into 128-byte sectors. When a line
in a sector is fetched, the hardware automatically fetches the other
line in the sector too. So from a false sharing perspective, the
effective line size is 128 bytes on the Netburst processors. (http://software.intel.com/en-us/forums/topic/292721)
IIRC, GCC has a __builtin_prefetch hint.
http://gcc.gnu.org/onlinedocs/gcc-3.3.6/gcc/Other-Builtins.html
has an excellent section on this. Basically, it suggests:
__builtin_prefetch (&array[i + LookAhead], rw, locality);
where rw is a 0 (prepare for read) or 1 (prepare for a write) value, and locality uses the number 0-3, where zero is no locality, and 3 is very strong locality.
Both are optional. LookAhead would be the number of elements to look ahead to. If memory access were 100 cycles, and the unrolled loops are two cycles apart, LookAhead could be set to 50 or 51.
There are two cases that need to be distinguished. Do you need to know the cache sizes at compile time or at runtime?
Determining the cache-size at compile-time
For some applications, you know the exact architecture that your code will run on, for example, if you can compile the code directly on the host machine. In that case, simplify looking up the size and hard-coding it is an option (could be automated in the build system). On most machines today, the L1 cache line should be 64 bytes.
If you want to avoid that complexity or if you need to support compilation on unknown architectures, you can use the C++17 feature std::hardware_constructive_interference_size as a good fallback. It will provide a compile-time estimation for the cache line, but be aware of its limitations. Note that the compiler cannot guess perfectly when it creates the binary, as the size of the cache-line is, in general, architecture dependent.
Determining the cache-size at runtime
At runtime, you have the advantage that you know the exact machine, but you will need platform specific code to read the information from the OS. A good starting point is the code snippet from this answer, which supports the major platforms (Windows, Linux, MacOS). In a similar fashion, you can also read the L2 cache size at runtime.
I would advise against trying to guess the cache line by running benchmarks at startup and measuring which one performed best. It might well work, but it is also error-prone if the CPU is used by other processes.
Combining both approaches
If you have to ship one binary and the machines that it will later run on features a range of different architectures with varying cache sizes, you could create specialized code parts for each cache size, and then dynamically (at application startup) choose the best fitting one.
The cache will usually do the right thing. The only real worry for normal programmer is false sharing, and you can't take care of that at runtime because it requires compiler directives.

mmap() vs. reading blocks

I'm working on a program that will be processing files that could potentially be 100GB or more in size. The files contain sets of variable length records. I've got a first implementation up and running and am now looking towards improving performance, particularly at doing I/O more efficiently since the input file gets scanned many times.
Is there a rule of thumb for using mmap() versus reading in blocks via C++'s fstream library? What I'd like to do is read large blocks from disk into a buffer, process complete records from the buffer, and then read more.
The mmap() code could potentially get very messy since mmap'd blocks need to lie on page sized boundaries (my understanding) and records could potentially lie across page boundaries. With fstreams, I can just seek to the start of a record and begin reading again, since we're not limited to reading blocks that lie on page sized boundaries.
How can I decide between these two options without actually writing up a complete implementation first? Any rules of thumb (e.g., mmap() is 2x faster) or simple tests?
I was trying to find the final word on mmap / read performance on Linux and I came across a nice post (link) on the Linux kernel mailing list. It's from 2000, so there have been many improvements to IO and virtual memory in the kernel since then, but it nicely explains the reason why mmap or read might be faster or slower.
A call to mmap has more overhead than read (just like epoll has more overhead than poll, which has more overhead than read). Changing virtual memory mappings is a quite expensive operation on some processors for the same reasons that switching between different processes is expensive.
The IO system can already use the disk cache, so if you read a file, you'll hit the cache or miss it no matter what method you use.
However,
Memory maps are generally faster for random access, especially if your access patterns are sparse and unpredictable.
Memory maps allow you to keep using pages from the cache until you are done. This means that if you use a file heavily for a long period of time, then close it and reopen it, the pages will still be cached. With read, your file may have been flushed from the cache ages ago. This does not apply if you use a file and immediately discard it. (If you try to mlock pages just to keep them in cache, you are trying to outsmart the disk cache and this kind of foolery rarely helps system performance).
Reading a file directly is very simple and fast.
The discussion of mmap/read reminds me of two other performance discussions:
Some Java programmers were shocked to discover that nonblocking I/O is often slower than blocking I/O, which made perfect sense if you know that nonblocking I/O requires making more syscalls.
Some other network programmers were shocked to learn that epoll is often slower than poll, which makes perfect sense if you know that managing epoll requires making more syscalls.
Conclusion: Use memory maps if you access data randomly, keep it around for a long time, or if you know you can share it with other processes (MAP_SHARED isn't very interesting if there is no actual sharing). Read files normally if you access data sequentially or discard it after reading. And if either method makes your program less complex, do that. For many real world cases there's no sure way to show one is faster without testing your actual application and NOT a benchmark.
(Sorry for necro'ing this question, but I was looking for an answer and this question kept coming up at the top of Google results.)
There are lots of good answers here already that cover many of the salient points, so I'll just add a couple of issues I didn't see addressed directly above. That is, this answer shouldn't be considered a comprehensive of the pros and cons, but rather an addendum to other answers here.
mmap seems like magic
Taking the case where the file is already fully cached1 as the baseline2, mmap might seem pretty much like magic:
mmap only requires 1 system call to (potentially) map the entire file, after which no more system calls are needed.
mmap doesn't require a copy of the file data from kernel to user-space.
mmap allows you to access the file "as memory", including processing it with whatever advanced tricks you can do against memory, such as compiler auto-vectorization, SIMD intrinsics, prefetching, optimized in-memory parsing routines, OpenMP, etc.
In the case that the file is already in the cache, it seems impossible to beat: you just directly access the kernel page cache as memory and it can't get faster than that.
Well, it can.
mmap is not actually magic because...
mmap still does per-page work
A primary hidden cost of mmap vs read(2) (which is really the comparable OS-level syscall for reading blocks) is that with mmap you'll need to do "some work" for every 4K page accessed in a new mapping, even though it might be hidden by the page-fault mechanism.
For a example a typical implementation that just mmaps the entire file will need to fault-in so 100 GB / 4K = 25 million faults to read a 100 GB file. Now, these will be minor faults, but 25 million page faults is still not going to be super fast. The cost of a minor fault is probably in the 100s of nanos in the best case.
mmap relies heavily on TLB performance
Now, you can pass MAP_POPULATE to mmap to tell it to set up all the page tables before returning, so there should be no page faults while accessing it. Now, this has the little problem that it also reads the entire file into RAM, which is going to blow up if you try to map a 100GB file - but let's ignore that for now3. The kernel needs to do per-page work to set up these page tables (shows up as kernel time). This ends up being a major cost in the mmap approach, and it's proportional to the file size (i.e., it doesn't get relatively less important as the file size grows)4.
Finally, even in user-space accessing such a mapping isn't exactly free (compared to large memory buffers not originating from a file-based mmap) - even once the page tables are set up, each access to a new page is going to, conceptually, incur a TLB miss. Since mmaping a file means using the page cache and its 4K pages, you again incur this cost 25 million times for a 100GB file.
Now, the actual cost of these TLB misses depends heavily on at least the following aspects of your hardware: (a) how many 4K TLB enties you have and how the rest of the translation caching works performs (b) how well hardware prefetch deals with with the TLB - e.g., can prefetch trigger a page walk? (c) how fast and how parallel the page walking hardware is. On modern high-end x86 Intel processors, the page walking hardware is in general very strong: there are at least 2 parallel page walkers, a page walk can occur concurrently with continued execution, and hardware prefetching can trigger a page walk. So the TLB impact on a streaming read load is fairly low - and such a load will often perform similarly regardless of the page size. Other hardware is usually much worse, however!
read() avoids these pitfalls
The read() syscall, which is what generally underlies the "block read" type calls offered e.g., in C, C++ and other languages has one primary disadvantage that everyone is well-aware of:
Every read() call of N bytes must copy N bytes from kernel to user space.
On the other hand, it avoids most the costs above - you don't need to map in 25 million 4K pages into user space. You can usually malloc a single buffer small buffer in user space, and re-use that repeatedly for all your read calls. On the kernel side, there is almost no issue with 4K pages or TLB misses because all of RAM is usually linearly mapped using a few very large pages (e.g., 1 GB pages on x86), so the underlying pages in the page cache are covered very efficiently in kernel space.
So basically you have the following comparison to determine which is faster for a single read of a large file:
Is the extra per-page work implied by the mmap approach more costly than the per-byte work of copying file contents from kernel to user space implied by using read()?
On many systems, they are actually approximately balanced. Note that each one scales with completely different attributes of the hardware and OS stack.
In particular, the mmap approach becomes relatively faster when:
The OS has fast minor-fault handling and especially minor-fault bulking optimizations such as fault-around.
The OS has a good MAP_POPULATE implementation which can efficiently process large maps in cases where, for example, the underlying pages are contiguous in physical memory.
The hardware has strong page translation performance, such as large TLBs, fast second level TLBs, fast and parallel page-walkers, good prefetch interaction with translation and so on.
... while the read() approach becomes relatively faster when:
The read() syscall has good copy performance. E.g., good copy_to_user performance on the kernel side.
The kernel has an efficient (relative to userland) way to map memory, e.g., using only a few large pages with hardware support.
The kernel has fast syscalls and a way to keep kernel TLB entries around across syscalls.
The hardware factors above vary wildly across different platforms, even within the same family (e.g., within x86 generations and especially market segments) and definitely across architectures (e.g., ARM vs x86 vs PPC).
The OS factors keep changing as well, with various improvements on both sides causing a large jump in the relative speed for one approach or the other. A recent list includes:
Addition of fault-around, described above, which really helps the mmap case without MAP_POPULATE.
Addition of fast-path copy_to_user methods in arch/x86/lib/copy_user_64.S, e.g., using REP MOVQ when it is fast, which really help the read() case.
Update after Spectre and Meltdown
The mitigations for the Spectre and Meltdown vulnerabilities considerably increased the cost of a system call. On the systems I've measured, the cost of a "do nothing" system call (which is an estimate of the pure overhead of the system call, apart from any actual work done by the call) went from about 100 ns on a typical modern Linux system to about 700 ns. Furthermore, depending on your system, the page-table isolation fix specifically for Meltdown can have additional downstream effects apart from the direct system call cost due to the need to reload TLB entries.
All of this is a relative disadvantage for read() based methods as compared to mmap based methods, since read() methods must make one system call for each "buffer size" worth of data. You can't arbitrarily increase the buffer size to amortize this cost since using large buffers usually performs worse since you exceed the L1 size and hence are constantly suffering cache misses.
On the other hand, with mmap, you can map in a large region of memory with MAP_POPULATE and the access it efficiently, at the cost of only a single system call.
1 This more-or-less also includes the case where the file wasn't fully cached to start with, but where the OS read-ahead is good enough to make it appear so (i.e., the page is usually cached by the time you want it). This is a subtle issue though because the way read-ahead works is often quite different between mmap and read calls, and can be further adjusted by "advise" calls as described in 2.
2 ... because if the file is not cached, your behavior is going to be completely dominated by IO concerns, including how sympathetic your access pattern is to the underlying hardware - and all your effort should be in ensuring such access is as sympathetic as possible, e.g. via use of madvise or fadvise calls (and whatever application level changes you can make to improve access patterns).
3 You could get around that, for example, by sequentially mmaping in windows of a smaller size, say 100 MB.
4 In fact, it turns out the MAP_POPULATE approach is (at least one some hardware/OS combination) only slightly faster than not using it, probably because the kernel is using faultaround - so the actual number of minor faults is reduced by a factor of 16 or so.
The main performance cost is going to be disk i/o. "mmap()" is certainly quicker than istream, but the difference might not be noticeable because the disk i/o will dominate your run-times.
I tried Ben Collins's code fragment (see above/below) to test his assertion that "mmap() is way faster" and found no measurable difference. See my comments on his answer.
I would certainly not recommend separately mmap'ing each record in turn unless your "records" are huge - that would be horribly slow, requiring 2 system calls for each record and possibly losing the page out of the disk-memory cache.....
In your case I think mmap(), istream and the low-level open()/read() calls will all be about the same. I would recommend mmap() in these cases:
There is random access (not sequential) within the file, AND
the whole thing fits comfortably in memory OR there is locality-of-reference within the file so that certain pages can be mapped in and other pages mapped out. That way the operating system uses the available RAM to maximum benefit.
OR if multiple processes are reading/working on the same file, then mmap() is fantastic because the processes all share the same physical pages.
(btw - I love mmap()/MapViewOfFile()).
mmap is way faster. You might write a simple benchmark to prove it to yourself:
char data[0x1000];
std::ifstream in("file.bin");
while (in)
{
in.read(data, 0x1000);
// do something with data
}
versus:
const int file_size=something;
const int page_size=0x1000;
int off=0;
void *data;
int fd = open("filename.bin", O_RDONLY);
while (off < file_size)
{
data = mmap(NULL, page_size, PROT_READ, 0, fd, off);
// do stuff with data
munmap(data, page_size);
off += page_size;
}
Clearly, I'm leaving out details (like how to determine when you reach the end of the file in the event that your file isn't a multiple of page_size, for instance), but it really shouldn't be much more complicated than this.
If you can, you might try to break up your data into multiple files that can be mmap()-ed in whole instead of in part (much simpler).
A couple of months ago I had a half-baked implementation of a sliding-window mmap()-ed stream class for boost_iostreams, but nobody cared and I got busy with other stuff. Most unfortunately, I deleted an archive of old unfinished projects a few weeks ago, and that was one of the victims :-(
Update: I should also add the caveat that this benchmark would look quite different in Windows because Microsoft implemented a nifty file cache that does most of what you would do with mmap in the first place. I.e., for frequently-accessed files, you could just do std::ifstream.read() and it would be as fast as mmap, because the file cache would have already done a memory-mapping for you, and it's transparent.
Final Update: Look, people: across a lot of different platform combinations of OS and standard libraries and disks and memory hierarchies, I can't say for certain that the system call mmap, viewed as a black box, will always always always be substantially faster than read. That wasn't exactly my intent, even if my words could be construed that way. Ultimately, my point was that memory-mapped i/o is generally faster than byte-based i/o; this is still true. If you find experimentally that there's no difference between the two, then the only explanation that seems reasonable to me is that your platform implements memory-mapping under the covers in a way that is advantageous to the performance of calls to read. The only way to be absolutely certain that you're using memory-mapped i/o in a portable way is to use mmap. If you don't care about portability and you can rely on the particular characteristics of your target platforms, then using read may be suitable without sacrificing measurably any performance.
Edit to clean up answer list:
#jbl:
the sliding window mmap sounds
interesting. Can you say a little more
about it?
Sure - I was writing a C++ library for Git (a libgit++, if you will), and I ran into a similar problem to this: I needed to be able to open large (very large) files and not have performance be a total dog (as it would be with std::fstream).
Boost::Iostreams already has a mapped_file Source, but the problem was that it was mmapping whole files, which limits you to 2^(wordsize). On 32-bit machines, 4GB isn't big enough. It's not unreasonable to expect to have .pack files in Git that become much larger than that, so I needed to read the file in chunks without resorting to regular file i/o. Under the covers of Boost::Iostreams, I implemented a Source, which is more or less another view of the interaction between std::streambuf and std::istream. You could also try a similar approach by just inheriting std::filebuf into a mapped_filebuf and similarly, inheriting std::fstream into a mapped_fstream. It's the interaction between the two that's difficult to get right. Boost::Iostreams has some of the work done for you, and it also provides hooks for filters and chains, so I thought it would be more useful to implement it that way.
I'm sorry Ben Collins lost his sliding windows mmap source code. That'd be nice to have in Boost.
Yes, mapping the file is much faster. You're essentially using the the OS virtual memory subsystem to associate memory-to-disk and vice versa. Think about it this way: if the OS kernel developers could make it faster they would. Because doing so makes just about everything faster: databases, boot times, program load times, et cetera.
The sliding window approach really isn't that difficult as multiple continguous pages can be mapped at once. So the size of the record doesn't matter so long as the largest of any single record will fit into memory. The important thing is managing the book-keeping.
If a record doesn't begin on a getpagesize() boundary, your mapping has to begin on the previous page. The length of the region mapped extends from the first byte of the record (rounded down if necessary to the nearest multiple of getpagesize()) to the last byte of the record (rounded up to the nearest multiple of getpagesize()). When you're finished processing a record, you can unmap() it, and move on to the next.
This all works just fine under Windows too using CreateFileMapping() and MapViewOfFile() (and GetSystemInfo() to get SYSTEM_INFO.dwAllocationGranularity --- not SYSTEM_INFO.dwPageSize).
mmap should be faster, but I don't know how much. It very much depends on your code. If you use mmap it's best to mmap the whole file at once, that will make you life a lot easier. One potential problem is that if your file is bigger than 4GB (or in practice the limit is lower, often 2GB) you will need a 64bit architecture. So if you're using a 32 environment, you probably don't want to use it.
Having said that, there may be a better route to improving performance. You said the input file gets scanned many times, if you can read it out in one pass and then be done with it, that could potentially be much faster.
Perhaps you should pre-process the files, so each record is in a separate file (or at least that each file is a mmap-able size).
Also could you do all of the processing steps for each record, before moving onto the next one? Maybe that would avoid some of the IO overhead?
I agree that mmap'd file I/O is going to be faster, but while your benchmarking the code, shouldn't the counter example be somewhat optimized?
Ben Collins wrote:
char data[0x1000];
std::ifstream in("file.bin");
while (in)
{
in.read(data, 0x1000);
// do something with data
}
I would suggest also trying:
char data[0x1000];
std::ifstream iifle( "file.bin");
std::istream in( ifile.rdbuf() );
while( in )
{
in.read( data, 0x1000);
// do something with data
}
And beyond that, you might also try making the buffer size the same size as one page of virtual memory, in case 0x1000 is not the size of one page of virtual memory on your machine... IMHO mmap'd file I/O still wins, but this should make things closer.
I remember mapping a huge file containing a tree structure into memory years ago. I was amazed by the speed compared to normal de-serialization which involves lot of work in memory, like allocating tree nodes and setting pointers.
So in fact I was comparing a single call to mmap (or its counterpart on Windows)
against many (MANY) calls to operator new and constructor calls.
For such kind of task, mmap is unbeatable compared to de-serialization.
Of course one should look into boosts relocatable pointer for this.
This sounds like a good use-case for multi-threading... I'd think you could pretty easily setup one thread to be reading data while the other(s) process it. That may be a way to dramatically increase the perceived performance. Just a thought.
To my mind, using mmap() "just" unburdens the developer from having to write their own caching code. In a simple "read through file eactly once" case, this isn't going to be hard (although as mlbrock points out you still save the memory copy into process space), but if you're going back and forth in the file or skipping bits and so forth, I believe the kernel developers have probably done a better job implementing caching than I can...
I think the greatest thing about mmap is potential for asynchronous reading with:
addr1 = NULL;
while( size_left > 0 ) {
r = min(MMAP_SIZE, size_left);
addr2 = mmap(NULL, r,
PROT_READ, MAP_FLAGS,
0, pos);
if (addr1 != NULL)
{
/* process mmap from prev cycle */
feed_data(ctx, addr1, MMAP_SIZE);
munmap(addr1, MMAP_SIZE);
}
addr1 = addr2;
size_left -= r;
pos += r;
}
feed_data(ctx, addr1, r);
munmap(addr1, r);
Problem is that I can't find the right MAP_FLAGS to give a hint that this memory should be synced from file asap.
I hope that MAP_POPULATE gives the right hint for mmap (i.e. it will not try to load all contents before return from call, but will do that in async. with feed_data). At least it gives better results with this flag even that manual states that it does nothing without MAP_PRIVATE since 2.6.23.