I want to understand how a certain worst case scenario of context switch happens. Say I have 10 CPU cores running a single process. Everything is CPU intensive, no thread is sleeping (waiting for I/O).
(I am mainly concerned with mainstream modern personal computer architectures and systems, typically x64 with Windows, Linux...)
Correct me if I'm wrong: running 10 CPU/RAM intensive independent threads is most often a near optimal situation. The amount of time spent in context switch is rather negligible. While the system may sometimes decide to re-attribute threads to different cores in a round-robin fashion causing a reset of RAM caches, it has a minor effect and works almost as if each thread was running on a single fixed core.
Only the main RAM bus may be a limitation since all threads share it, but it's not the point I'm interested in here. Reducing the number of threads will not increase the throughput anyway.
Now assume you still have 10 cores but run 1000 threads. The scheduler could theoretically decide to switch rarely (say every second) running 10 threads for a second, then 10 others... and the whole thing would still be close to optimal performance (throughput).
But it does not seem to be the case and it looks like threads are switched intensively causing a strongly suboptimal performance (throughput). Am I right about it? What is the main cause for this suboptimal performance? A few numbers would be nice if you have any idea of orders of magnitude of (for example): switches per second, performance loss caused by switching...
I'm going to answer my own question (after some search).
On windows, the number of context switches can be measured with performance counters: https://technet.microsoft.com/en-us/library/cc938606.aspx
I measured it on my machine (core i7/Windows 10) and the order of magnitude is around 1000/s by core when the number of running threads is more than the number of cores (and these threads are full CPU).
The time needed for a context switch varies quite a bit depending on:
what registers need to be saved
if FPU registers need to be saved
the processor model (of course)
You can read: https://www.quora.com/How-long-does-a-context-switch-take or http://blog.tsunanet.net/2010/11/how-long-does-it-take-to-make-context.html
A slightly pessimistic avg. order of magnitude seems to be 1000 ns. Thus the total time for all context switches on each core is 1ms per second, that is 0.1%.
This does not depend on the number of threads: if you run 100 or 1000 threads, the number of switches does not change. As a conclusion the time spent in context switching is somehow negligible.
This reasoning is correct as long as the threads are pure CPU with only small memory read/write like a few local variables. I ran a test with full CPU threads and the difference between a few and 1000 threads is not noticeable.
But the situation changes when RAM is involved and switches makes CPU (memory) cache less efficient. A worse case is when:
computation can be split into 1000 independent "data" parts
each part of the data fits just into the memory cache (say L1 or L2) of a core
each part needs to be read many times
In this situation, running 10 threads to completion, then ten others... would take full advantage of the cache, while running 1000 threads at a time would causes the cache to be useful only during 1ms.
But if the data of several threads could fit into the cache, or if the threads read common data to some degree, or if each thread reads the data just once, then it is possible that running 1000 threads vs. running 10 threads a hundred times will have similar throughput.
It is more a matter a adapting parallelism to memory access. And it depends very much on the way memory needs to be accessed.
The time spend in context switching is negligible, the time lost because of wrong usage of caches may sometimes be problem, sometimes not, depending on how the memory is accessed and shared.
Related
I am observing strange effects with the CPU percentage as shown in e.g. top or htop on Linux (Ubuntu 16.04) for one special application. The application uses many threads (around 1000). Each thread has one computational task. About half of these tasks need to be computed once per "trigger" - the trigger is an external event received exactly every 100ms. The other threads are mostly sleeping (waiting for user interaction) and hence do not play a big role here. So to summarise: many threads are waking up basically simultaneously within a short period of time, doing there (relatively short) computation and going back to sleep again.
Since the machine running this application has 8 virtual CPUs (4 cores each 2 threads, it's an i7-3612QE), only 8 threads can really wake up at a time, so many threads will have to wait. Also some of these tasks have interdependencies, so they anyway have to wait, but I think as an approximation one can think of this application as a bunch of threads going to the runnable state at the same time every 100ms and each doing only a short computation (way below 1ms of CPU time each).
Now coming to the strange effect: If I look at the CPU percentage in "top", it shows something like 250%. As far as I know, top looks on the CPU time (user + system) the kernel accounts for this process, so 250% would mean the process uses 3 virtual CPUs on average. So far so good. Now, if I use taskset to force the entire process to use only a single virtual CPU, the CPU percentage drops to 80%. The application has internal accounting which tells me that still all data is being processed. So the application is doing the same amount of work, but it seemingly uses less CPU resources. How can that be? Can I really trust the kernel CPU time accounting, or is this an artefact of the measurement?
The CPU percentage also goes down, if I start other processes which take a lot of CPU, even if the do nothing ("while(true);") and are running at low priority (nice). If I launch 8 of these CPU-eating processes, the application reaches again 80%. With fewer CPU-eaters, I get gradually higher CPU%.
Not sure if this plays a role: I have used the profiler vtune, which tells me my application is actually quite inefficient (only about 1 IPC), mostly because it's memory bound. This does not change if I restrict the process to a single virtual CPU, so I assume the effect is not caused by a huge increase in efficiency when running everything on the same core (which would be strange anyway).
My question was essentially already answered by myself in the last paragraph: The process is memory bound. Hence not the CPU is the limited resource but the memory bandwidth. Allowing such process to run on multiple CPU cores in parallel will mainly have the effect that more CPU cores are waiting for data to arrive from RAM. This is counted as CPU load, since the CPU is executing the thread, but just quite slowly. All my other observations go along with this.
I was doing some testing with multi-threading on a linux virtual machine, and I implemented a benchmark with 10 threads (in this application each instruction would be executed 10x times more than in the single-thread scenario) and i was tweaking with the number of "physical cores" from the VM settings and with the single thread case I obtain 3s on average independently of the number of physical cores, If the number of cores is set to 1, and I run the multi-thread version, the execution time will be 30s. If I run it with 2 cores I obtain 15s and with 8 cores (the maximum number I can set) I obtain 6s, I obtain this dependancy due to the fact that I´m executing 10x times each instruction or is always like this?
If you have N threads running on N cores, and if they are all doing pure computation (i.e., not waiting for any I/O devices), and if they are all completely independent of each other, then they should be able to do N times as much work in a given amount of time as a single thread can do in the same amount of time.
But, that's if they are completely independent. That's a hard thing to achieve. For example, if the threads can't each do all of their work in their own, independent cache (e.g., in L1 cache,) then they will compete with each other for access to the main memory. They will sometimes have to wait for one another, because only one core can access main memory at any given moment. So, if the threads need to use memory, then the speedup will be somewhat less than N times.
If the threads need to share data in main memory, then it gets worse because then they will need to use mutual exclusion locks. One thread may keep a lock locked while it executes dozens of instructions, and any other thread that wants the same lock will have to wait until it is finished.
If the threads need to synchronize with each other/communicate with each other, then it gets worse still because unless their work loads are carefully balanced, a thread with less work to do may spend long periods of time awaiting signals from threads that have more work to do.
It's not unusual for a novice programmer to invent a multi-threaded version of some single-threaded algorithm, and find out that the multi-threaded version actually is slower than the single-threaded version.
There are some algorithms, for which even an expert programmer can't get much speed up by throwing more threads at it.
The following is an excerpt from the book Java Concurrency in Practice, Chapter 12.2 Testing for Performance where the author talks about throughput of a bounded buffer implementation.
Figure 12.1 shows some sample results on a 4-way machine, using buffer
capacities of 1, 10, 100, and 1000. We see immediately that a buffer
size of one causes very poor throughput; this is because each thread
can make only a tiny bit of progress before blocking and waiting for
another thread. Increasing buffer size to ten helps dramatically, but
increases past ten offer diminishing returns.
It may be somewhat puzzling at first that adding a lot more threads
degrades performance only slightly. The reason is hard to see from the
data, but easy to see on a CPU performance meter such as perfbar while
the test is running: even with many threads, not much computation is
going on, and most of it is spent blocking and unblocking threads. So
there is plenty of CPU slack for more threads to do the same thing
without hurting performance very much.
However, be careful about concluding from this data that you can
always add more threads to a producer-consumer program that uses a
bounded buffer. This test is fairly artificial in how it simulates the
application; the producers do almost no work to generate the item
placed on the queue, and the consumers do almost no work with the item
retrieved. If the worker threads in a real producer-consumer
application do some nontrivial work to produce and consume items (as
is generally the case), then this slack would disappear and the
effects of having too many threads could be very noticeable. The
primary purpose of this test is to measure what constraints the
producer-consumer handoff via the bounded buffer imposes on overall
throughput.
What does the author mean by cpu slack here? Why will the throughput degrade not degrade more and more as more number of threads are being added? I am not following the reasoning given by the author regarding the slight degradation of performance while adding more and more threads , assuming that the bound on the buffer size is kept constant.
Edit: I can think of one reason :since in this case no real work is being done by threads , so the classic problem of increased traffic on shared memory bus, number of cache misses due to context switching of threads are not playing a major role as more and more threads are being added. The situation is going to change once the threads start doing some more work. Is that what the author is trying to convey here in the third paragraph?
There is no formal term such as CPU slack. The author simply means that the CPU is not fully utilised in doing meaningful work because most time is spent waiting to successfully acquire a mutually exclusive lock. The author is calling the unused capacity of the CPU, the CPU slack.
NOTE: The associated code tests a multiple producer / multiple consumer scenario, with an equal number of producers and consumers.
EDIT: In the later discussion they talk about the effect of adding more threads if a) the threads do almost no work, and b) the threads occupy the CPU substantially for every produced or consumed item. I will try to explain the difference with some slightly artificial scenarios.
Suppose that locking takes 1 time unit actively, and 8 time units passively by waiting. Passive waiting does not occupy the CPU.
Case 1: Producer-Consumer cost is 1 time unit.
So we currently account for 2 time units of CPU time, with an
additional 8 time units of passive waiting time. So we have 8/10
available CPU time units.
If we now want to double the number of threads, we need to accommodate
an additional 2 time units (1 for producer-consumer stuff, and 1 for
active locking time). That would eat into our supply of available CPU
time -- but we have enough.
Case 2: Producer-Consumer cost is 11 time units.
So we currently account for 11+1=12 time units of CPU time, with an additional 8 time units of passive waiting time. So we have 8/20 available CPU time units.
If we now want to double the number of threads, we need to accommodate an additional 12 time units (11 for producer-consumer stuff, and 1 for active locking time). That goes beyond the available CPU time units. Something has to give -- so waiting time will increase, and throughput will suffer.
So in case 2, the amount of real work reduces the amount of time available for new threads, thereby increasing the observed effect of locking contention on the throughput. It would have been nice if they had also included figures for this imagined scenarios in the book. It would have made their hand-wavy argument easier to follow.
I think cpu slack is the resource. According to Wikipedia, it is referred to the amount of time left after a job if the job was started now.
Plenty of cpu slack means much computation resources. When Consumer/Producer do something nontrivial, cpu slack decreases and impacts throughput.
Lets assume we have fixed amount of calculation work, without blocking, sleeping, i/o-waiting. The work can be parallelized very well - it consists of 100M small and independent calculation tasks.
What is faster for 4-core CPU - to run 4 threads or... lets say 50? Why second variant should be slover and how much slover?
As i assume: when you run 4 heavy threads on 4-core CPU without another CPU-consuming processes/threads, scheduler is allowed to not move the threads between cores at all; it has no reason to do that in this situation. Core0 (main CPU) will be responsible for executing interruption handler for hardware timer 250 times per second (basic Linux configuration) and other hardware interruption handlers, but another cores may not feel any worries.
What is the cost of context switching? The time for store and restore CPU registers for different context? What about caches, pipelines and various code-prediction things inside CPU? Can we say that each time we switch context, we hurt caches, pipelines and some code-decoding facilities in CPU? So more threads executing on a single core, less work they can do together in comparison to their serial execution?
Question about caches and another hardware optimization in multithreading environment is the interesting question for me now.
As #Baile mentions in the comments, this is highly application, system, environment-specific.
And as such, I'm not going to take the hard-line approach of mentioning exactly 1 thread for each core. (or 2 threads/core in the case of Hyperthreading)
As an experienced shared-memory programmer, I have seen from my experience that the optimal # of threads (for a 4 core machine) can range anywhere from 1 to 64+.
Now I will enumerate the situations that can cause this range:
Optimal Threads < # of Cores
In certain tasks that are very fine-grained paralleled (such as small FFTs), the overhead of threading is the dominant performance factor. In some cases, it's it not helpful to parallelize at all. In some cases, you get speedup with 2 threads, but backwards scaling at 4 threads.
Another issue is resource contention. Even if you have a highly parallelizable task that can easily split across 4 cores/threads, you may be bottlenecked by memory bandwidth and cache effects. So often, you find that 2 threads will be just as fast as 4 threads. (as if often the case with very large FFTs)
Optimal Threads = # of Cores
This is the optimal case. No need to explain here - one thread per core. Most embarrassingly parallel applications that are not memory or I/O bound fit right here.
Optimal Threads > # of Cores
This is where it gets interesting... very interesting. Have you heard about load-imbalance? How about over-decomposition and work-stealing?
Many parallelizable applications are irregular - meaning that the tasks do not split into sub-tasks of equal size. So if you may end up splitting a large task into 4 unequal sizes, assign them to 4 threads and run them on 4 cores... the result? Poor parallel performance because 1 thread happened to get 10x more work than the other threads.
A common solution here is to over-decompose the task into many sub-tasks. You can either create threads for each one of them (so now you get threads >> cores). Or you can use some sort of task-scheduler with a fixed number of threads. Not all tasks are suited for both, so quite often, the approach of over-decomposing a task to 8 or 16 threads for a 4-core machine gives optimal results.
Although spawning more threads can lead to better load-balance, the overhead builds up. So there's typically an optimal point somewhere. I've seen as high as 64 threads on 4 cores. But as mentioned, it's highly application specific. And you need to experiment.
EDIT : Expanding answer to more directly answer the question...
What is the cost of context switching? The time for store and restore
CPU registers for different context?
This is very dependent on the environment - and is somewhat difficult to measure directly. Short answer: Very Expensive This might be a good read.
What about caches, pipelines and various code-prediction things inside
CPU? Can we say that each time we switch context, we hurt caches,
pipelines and some code-decoding facilities in CPU?
Short answer: Yes When you context switch out, you likely flush your pipeline and mess up all the predictors. Same with caches. The new thread is likely to replace the cache with new data.
There's a catch though. In some applications where the threads share the same data, it's possible that one thread could potentially "warm" the cache for another incoming thread or another thread on a different core sharing the same cache. (Although rare, I've seen this happen before on one of my NUMA machines - superlinear speedup: 17.6x across 16 cores!?!?!)
So more threads executing on a single core, less work they can do
together in comparison to their serial execution?
Depends, depends... Hyperthreading aside, there will definitely be overhead. But I've read a paper where someone used a second thread to prefetch for the main thread... Yes it's crazy...
Creating 50 threads will actually hurt performance, not improve it. It just doesn't make any sense.
Ideally you should make the 4 threads, not more, not less. There will be some overhead because of context switching, but that is unavoidable. The OS/services/other applications threads should too be executed. But nowadays with so powerful and lighting-fast CPUs this is of no concern since those OS threads will only take less that 2 % of the CPU's time. Almost all of them will be in blocked state while your program is running.
You might think that, since performance is of critical importance, you should code those small critical areas in low-level assembly language. Modern programming languages allow this.
But seriously... compilers and, in case of Java, the JVM, will optimize those portions so well that it just isn't worth it (unless you actually want to exercise something like this). Instead of your calculations finishing in 100 seconds, they'll finish in 97 or 98. The question you must ask yourself is: is it worth all those hours of coding and debugging ?
You asked about the time cost of context switching. These days, these are extremely low. Look at modern day dual-core CPUs that run Windows 7 for example. If you start an Apache web server on that machine and a MySQL database server, you will easily go over 800 threads. The machine just doesn't feel it. To see how low this cost is, read here: How to estimate the thread context switching overhead? . To spare you the searching/reading part: context switching can be done hundreds of thousands of times per second.
4 threads are faster if you can program your 40 tasks switching better than Operating System does.
If you can use 4 threads, use them. There's no way 50 will go faster than 4 on a 4-core machine. All you get is more overhead.
Of course, you're describing an ideal non-real-world situation, so whatever you are actually building, you'll need to measure in order to understand how the performance is affected.
There is Hyperthreading technology which can handle more that one thread per CPU, but it is hardly dependent on type of calculation you want to do. Consider using of GPU or very low assembly language to achieve maximum power.
If you're spawning multiple threads (or processes) concurrently, is it better to spawn as many as the number of physical processors or the number of logical processors, assuming the task is CPU-bound? Or is it better to do something in between (say, 3 threads)?
Does the performance depend on the kind of instructions that are getting executed (say, would non-local memory access be much different from cache hits)? If so, in which cases is it better to take advantage of hyperthreading?
Update:
The reason I'm asking is, I remember reading somewhere that if you have as many tasks as the number of virtual processors, tasks on the same physical core can sometimes starve some CPU resources and prevent each other from getting as many resources as needed, possibly decreasing performance. That's why I'm wondering if having as many threads as virtual cores is a good idea.
The performance depends on a huge variety of factors. Most tasks are not strictly CPU bound, since even if all of the data is in memory it is usually not on-board in the processor cache. I have seen examples (like this one) where memory access patterns can dramatically change the performance profile of a given 'parallel' process.
In short, there is no perfect number for all situations.
Chances are pretty good that you will see a performance improvement running 2 threads per core with HyperThreading enabled. Jobs that appear to be entirely CPU bound usually aren't, and HyperThreading can extract a few "extra" cycles out of the occasional interrupt or context switch.
On the other hand, with a core iX processor that has Turbo Boost, you might actually do better running 1 thread per core to encourage the CPU to overclock itself.
At work, we routinely run many-core servers at full CPU doing various kinds of calculation for days at a time. A while back we measured the performance difference with and without HT. We found that on average, with HyperThreading, and running twice as many jobs at once, we could complete the same amount of jobs about 10% faster than than without HyperThreading.
Assume that 2 × cores is a good place to start, but the bottom line is: measure!
I remember info that hyperthreading can give you up to 30% of performance boost. in general you'd better to treat them as 4 different cores. of course in some specific circumstances (e.g. having the same long running task bound to each core) you can divide your processing better taking into account that some cores are just logical ones
more info about hyperthreading itself here
Using Hyperthreading to run two threads on the same core, when both threads have similar memory access patterns but access disjoint data structures, would be very roughly equivalent to running them on two separate cores each with half the cache. If the memory-access patterns are such that half the cache would be sufficient to prevent thrashing, performance may be good. If the memory-access patterns are such that halving the cache induces thrashing, there may be a ten-fold performance hit (implying one would have been much better off without hyperthreading).
On the other hand, there are some situations where hyperthreading may be a huge win. If many threads will all be reading and writing the same shared data using lock-free data structures, and all threads must see a consistent view of the data, trying to run threads on disjoint processor may cause thrashing since only one processor at a time may have read-write access to any given cache line; running such a threads on two cores may take longer than running only one at a time. Such cache arbitration is not required, however, when a piece of data is accessed by multiple threads on a single core. In those cases, hyperthreading can be a huge win.
Unfortunately, I don't know any way to give the scheduler any "hints" to suggest that some threads should share a core when possible, while others should run separately when possible.
HT allows a boost of approximately 10-30% for mostly cpu-bound tasks that use the extra virtual cores. Although these tasks may seem CPU-bound, unless they are custom made assembly, they will usually suffer from IO waits between RAM and local cache. This allows one thread running on a physical HT-enabled core to work while the other thread is waiting for IO. This does come with a disadvantage though, as two threads share the same cache/bus, which will result in less resources each which may cause both threads to pause while waiting for IO.
In the last case, running a single thread will decrease the maximum simultaneous theoretical processing power(by 10-30%) in favor of running a single thread without the slowdown of cache thrashing which may be very significant in some applications.
Choosing which cores to use is just as important as choosing how many threads to run. If each thread is CPU-bound for roughly the same duration it is best to set the affinity such that threads using mostly different resources find themselves on different physical cores and threads using common resources be grouped to the same physical cores(different virtual core) so that common resources can be used from the same cache without extra IO wait.
Since each program has different CPU-usage characteristics and cache thrashing may or may not be a major slowdown(it usually is) it is impossible to determine what the ideal number of threads should be without profiling first. One last thing to note is that the OS/Kernel will also require some CPU and cache space. It is usually ideal to keep a single (physical)core set aside for the OS if real-time latency is required on CPU-bound threads so as to avoid sharing cache/cpu resources. If threads are often waiting for IO and cache thrashing is not an issue, or if running a real-time OS specifically designed for the application, you can skip this last step.
http://en.wikipedia.org/wiki/Thrashing_(computer_science)
http://en.wikipedia.org/wiki/Processor_affinity
All of the other answers already give lots of excellent info. But, one more point to consider is that the SIMD unit is shared between logical cores on the same die. So, if you are running threads with SSE code, do you run them on all 4 logical cores, or just spawn 2 threads (assuming you have two chips)? For this odd case, best to profile with your app.