Related
Given:
A fully CPU bound very large (i.e. more than a few CPU cycles) job, and
A CPU with 4 physical and total 8 logical cores,
is it possible that 8, 16 and 28 threads perform better than 4 threads? My understanding is that 4 threads would have lesser context switches to perform and will have lesser overhead in any sense than 8, 16 or 28 threads would have on a 4 physical core machine. However, the timings are -
Threads Time Taken (in seconds)
4 78.82
8 48.58
16 51.35
28 52.10
The code used to test get the timings is mentioned in the Original Question section below. The CPU specifications are also given at the bottom.
After reading the answers that various users have provided and information given in the comments, I am able to finally boil down the question to what I wrote above. If the question above gives you the complete context, you can skip the original question below.
Original Question
What does it mean when we say
Hyper-threading works by duplicating certain sections of the
processor—those that store the architectural state—but not duplicating
the main execution resources. This allows a hyper-threading processor
to appear as the usual "physical" processor and an extra "logical"
processor to the host operating system
?
This question is asked on SO today and it basically tests the performance of multiple threads doing the same work. It has the following code:
private static void Main(string[] args)
{
int threadCount;
if (args == null || args.Length < 1 || !int.TryParse(args[0], out threadCount))
threadCount = Environment.ProcessorCount;
int load;
if (args == null || args.Length < 2 || !int.TryParse(args[1], out load))
load = 1;
Console.WriteLine("ThreadCount:{0} Load:{1}", threadCount, load);
List<Thread> threads = new List<Thread>();
for (int i = 0; i < threadCount; i++)
{
int i1 = i;
threads.Add(new Thread(() => DoWork(i1, threadCount, load)));
}
var timer = Stopwatch.StartNew();
foreach (var thread in threads) thread.Start();
foreach (var thread in threads) thread.Join();
timer.Stop();
Console.WriteLine("Time:{0} seconds", timer.ElapsedMilliseconds/1000.0);
}
static void DoWork(int seed, int threadCount, int load)
{
var mtx = new double[3,3];
for (var i = 0; i < ((10000000 * load)/threadCount); i++)
{
mtx = new double[3,3];
for (int k = 0; k < 3; k++)
for (int l = 0; l < 3; l++)
mtx[k, l] = Math.Sin(j + (k*3) + l + seed);
}
}
(I have cut out a few braces to bring the code in a single page for quick readability.)
I ran this code on my machine for replicating the issue. My machine has 4 physical cores and 8 logical ones. The method DoWork() in the code above is completely CPU bound. I felt that hyper-threading could contribute to maybe a 30% speedup (because here we have as many CPU bound threads as the physical cores (i.e. 4)). But it nearly does attain 64% performance gain. When I ran this code for 4 threads, it took about 82 seconds and when I ran this code for 8, 16 and 28 threads, it ran in all the cases in about 50 seconds.
To summarize the timings:
Threads Time Taken (in seconds)
4 78.82
8 48.58
16 51.35
28 52.10
I could see that CPU usage was ~50% with 4 threads. Shouldn't it be ~100%? After all my processor has only 4 physical cores. And the CPU usage was ~100% for 8 and 16 threads.
If somebody can explain the quoted text at the start, I hope to understand hyperthreading better with it and in turn hope to get the answer to Why would a fully CPU bound process work better with hyperthreading?.
For the sake of completion,
I have Intel Core i7-4770 CPU # 3.40 GHz, 3401 MHz, 4 Core(s), 8 Logical Processor(s).
I ran the code in Release mode.
I know that the way timings are measured is bad. This will only give the time for slowest thread. I took the code as it is from the other question. However, what is the justification for 50% CPU usage when running 4 CPU bound threads on a 4 physical core machine?
CPU pipeline
Each instruction has to go through several steps in the pipeline to be fully executed. At the very least, it must be decoded, sent to execution unit, then actually executed there. There are several execution units on modern CPUs, and they can execute instructions completely in parallel. By the way, the execution units are not interchangeable: some operations can only be done on a single execution unit. For example, memory loads are usually specialized to one or two units, memory stores are exclusively sent to another unit, all the calculations are done by some other units.
Knowing about the pipeline, we may wonder: how can CPU work so fast, if we write purely sequental code and each instruction has to go through so many pipeline stages? Here is the answer: processor executes instructions in out-of-order fashion. It has a large reorder buffer (e.g. for 200 instructions), and it pushes many instructions through its pipeline in parallel. If at any moment some instruction cannot be executed for any reason (waits for data from slow memory, depends on other instruction not yet finished, whatever), then it is delayed for some cycles. During this time processor executes some new instructions, which are located after the delayed instructions in our code, given that they do not depend on the delayed instructions in any way.
Now we can see the problem of latency. Even if an instruction is decoded and all of its inputs are already available, it would take it several cycles to be executed completely. This delay is called instruction latency. However, we know that at this moment processor can execute many other independent instructions, if there are any.
If an instruction loads data from L2 cache, it has to wait about 10 cycles for the data to be loaded. If the data is located only in RAM, then it would take hundreds of cycles to load it to processor. In this case we can say that the instruction has high latency. It is important for maximum performance to have some other independent operations to execute at this moment. This is sometimes called latency hiding.
At the very end, we have to admit that most of real code is sequental in its nature. It has some independent instructions to execute in parallel, but not too many. Having no instructions to execute causes pipeline bubbles, and it leads to inefficient usage of processor's transistors. On the other hand, instructions of two different threads are automatically independent in almost all cases. This leads us directly to the idea of hyper-threading.
P.S. You might want to read Agner Fog's manual to better understand internals of modern CPUs.
Hyper-threading
When two threads are executed in hyper-threading mode on a single core, the processor can interleave their instructions, allowing to fill bubbles from the first thread with instructions of the second thread. This allows to better utilize processor's resources, especially in case of ordinary programs. Note that HT may help not only when you have a lot of memory accesses, but also in heavily sequental code. A well-optimized computational code may fully utilize all resources of CPU, in which case you will see no profit from HT (e.g. dgemm routine from well-optimized BLAS).
P.S. You might want to read Intel's detailed explanation of hyper-threading, including info about which resources are duplicated or shared, and discussion about performance.
Context switches
The context is an internal state of CPU, which at least includes all the registers. When execution thread changes, OS has to do a context switch (detailed description here). According to this answer, context switch takes about 10 microseconds, while the time quant of scheduler is 10 milliseconds or more (see here). So context switches do not affect total time much, because they are done seldom enough. Note that competition for CPU caches between threads can increase the effective cost of switches in some cases.
However, in case of hyper-threading each core has two states internally: two sets of registers, shared caches, one set of execution units. As a result, the OS has no need to do any context switches when you run 8 threads on 4 physical cores. When you run 16 threads on quad-core, the context switches are performed, but they take small part of the overall time, as explained above.
Process manager
Speaking of CPU utilization that you see in the process manager, it does not measure the internals of CPU pipeline. Windows can only notice when a thread returns execution to OS in order to: sleep, wait for mutex, wait for HDD, and do other slow things. As a result, it thinks that a core is fully used if there is a thread working on it, which does not sleep or wait for anything. For instance, you may check that running endless loop while (true) {} leads to full utilization of CPU.
I could see that CPU usage was ~50% with 4 threads. Shouldn't it be ~100%?
No, it shouldn't.
what is the justification for 50% CPU usage when running 4 CPU bound threads on a 4 physical core machine?
This is simply how CPU utilization is reported in Windows (and on at least some other OS's too, by the way). A HT CPU shows up as two cores to the operating system, and is reported as such.
Thus, Windows sees an eight-core machine, when you have four HT CPUs. You'll see eight different CPU graphs if you look at the "Performance" tab in Task Manager, and the total CPU utilization is computed with 100% utilization being the full utilization of these eight cores.
If you are only using four threads, then these threads cannot fully utilize the available CPU resources and that explains the timings. They can, at most, use four of the eight cores available and so of course your utilization will max out at 50%. Once you go past the number of logical cores (8), runtime increases again; you are adding scheduling overhead without adding any new computational resources in that case.
By the way…
HyperThreading has improved quite a lot from the old days of shared cache and other limitations, but it will still never provide the same throughput benefit that a full CPU could, as there remains some contention within the CPU. So even ignoring OS overhead, your 35% improvement in speed seems pretty good to me. I often see no more than a 20% speed-up adding the extra HT cores to a computationally-bottlenecked process.
I can't explain the sheer volume of speed-up that you observed: 100% seems way too much of an improvement for Hyperthreading. But I can explain the principles in place.
The main benefit to Hyperthreading is when a processor has to switch between threads. Whenever there are more threads than there are CPU cores (true 99.9997% of the time) and the OS decides to switch to a different thread, it has to perform (most of) the following steps:
Save the state of the current thread: this includes the stack, the state of the registers, and the program counter. where they get saved depends on the architecture, but generally speaking they'll either get saved in cache or in memory. Either way, this step takes time.
Put the Thread into "Ready" state (as opposed to "Running" state).
Load the state of the next thread: again, including the stack, the registers, and the program counter, which once again, is a step that takes time.
Flip the Thread into "Running" state.
In a normal (non-HT) CPU, the number of cores it has is the quantity of processing units. Each of these contain registers, program counters (registers), stack counters (registers), (usually) individual cache, and complete processing units. So if a normal CPU has 4 cores, it can run 4 threads simultaneously. When a thread is done (or the OS has decided that it's taking too much time and needs to wait its turn to start again), the CPU needs to follow those four steps to unload the thread and load in the new one before execution of the new one can begin.
In a HyperThreading CPU, on the other hand, the above holds true, but in addition, Each core has a duplicated set of Registers, Program Counters, Stack Counters, and (sometimes) cache. What this means is that a 4-core CPU can still only have 4 threads running simultaneously, but the CPU can have "preloaded" threads on the duplicated registers. So 4 threads are running, but 8 threads are loaded onto the CPU, 4 active, 4 inactive. Then, when it's time for the CPU to switch threads, instead of having to perform the loading/unloading at the moment the threads need to switch out, it simply "toggles" which thread is active, and performs the unloading/loading in the background on the newly "inactive" registers. Remember the two steps I suffixed with "these steps take time"? In a Hyperthreaded system, steps 2 and 4 are the only ones that need to be performed in real-time, whereas steps 1 and 3 are performed in the background in the hardware (divorced from any concept of threads or processes or CPU cores).
Now, this process doesn't completely speed up multithreaded software, but in an environment where threads often have extremely small workloads that they perform very frequently, the quantity of thread-switches can be expensive. Even in environments that don't conform to that paradigm, there can be benefits from Hyperthreading.
Let me know if you need any clarifications. It's been a few years since CS250, so I may be mixing up terminology here-or-there; let me know if I'm using the wrong terms for something. I'm 99.9997% certain that everything I'm describing is accurate in terms of the logic of how it all works.
Hyper-threading works by interleaving instructions in the processor execution pipeline. While the processor is performing read-write operations on one 'thread' it is performing logical evaluation on the other 'thread', keeping them separate and giving you a perceived doubling in performance.
The reason you get such a big speedup is because there is no branching logic in your DoWork method. It is all a big loop with a very predictable execution sequence.
A processor execution pipeline has to go through several clock cycles to execute a single calculation. The processor attempts to optimise the performance by pre-loading the execution buffer with the next few instructions. If the instruction loaded is actually a conditional jump (such as an if statement), this is bad news, because the processor has to flush the entire pipeline and fetch instructions from a different part of memory.
You may find that if you put if statements in your DoWork method, you will not get 100% speedup...
Recently I've been analyzing how my parallel computations actually speed up on 16-core processor. And the general formula that I concluded - the more threads you have the less speed per core you get - is embarassing me. Here are the diagrams of my cpu load and processing speed:
So, you can see that processor load increases, but speed increases much slower. I want to know why such an effect takes place and how to get the reason of unscalable behaviour.
I've made sure to use Server GC mode.
I've made sure that I'm parallelizing appropriate code as soon as code does nothing more than
Loads data from RAM (server has 96 GB of RAM, swap file shouldn't be hit)
Performs not complex calculations
Stores data in RAM
I've profiled my application carefully and found no bottlenecks - looks like each operation becomes slower as thread number grows.
I'm stuck, what's wrong with my scenario?
I use .Net 4 Task Parallel Library.
You will always get this kind of curve, it's called Amdahl's law.
The question is how soon it will level off.
You say you checked your code for bottlenecks, let's assume that's correct. Then there is still the memory bandwidth and other hardware factors.
The key to a linear scalability - in the context of where going from one to two cores doubles the throughput - is to use shared resources as little as possible. This means:
don't use hyperthreading (because the two threads share the same core resource)
tie every thread to a specific core (otherwise the OS will juggle the
threads between cores)
don't use more threads than there are cores (the OS will swap in and
out)
stay inside the core's own caches - nowadays the L1 & L2 caches
don't venture into the L3 cache or RAM unless it is absolutely
necessary
minimize/economize on critical section/synchronization usage
If you've come this far you've probably profiled and hand-tuned your code too.
Thread pools are a compromise and not suited for uncompromising, high-performance applications. Total thread control is.
Don't worry about the OS scheduler. If your application is CPU-bound with long computations that mostly does local L1 & L2 memory accesses it's a better performance bet to tie each thread to its own core. Sure the OS will come in but compared to the work being performed by your threads the OS work is negligible.
Also I should say that my threading experience is mostly from Windows NT-engine machines.
_______EDIT_______
Not all memory accesses have to do with data reads and writes (see comment above). An often overlooked memory access is that of fetching code to be executed. So my statement about staying inside the core's own caches implies making sure that ALL necessary data AND code reside in these caches. Remember also that even quite simple OO code may generate hidden calls to library routines. In this respect (the code generation department), OO and interpreted code is a lot less WYSIWYG than perhaps C (generally WYSIWYG) or, of course, assembly (totally WYSIWYG).
A general decrease in return with more threads could indicate some kind of bottle neck.
Are there ANY shared resources, like a collection or queue or something or are you using some external functions that might be dependent on some limited resource?
The sharp break at 8 threads is interesting and in my comment I asked if the CPU is a true 16 core or an 8 core with hyper threading, where each core appears as 2 cores to the OS.
If it is hyper threading, you either have so much work that the hyper threading cannot double the performance of the core, or the memory pipe to the core cannot handle twice the data through put.
Are the work performed by the threads even or are some threads doing more than others, that could also indicate resource starvation.
Since your added that threads query for data very often, that indicates a very large risk of waiting.
Is there any way to let the threads get more data each time? Like reading 10 items instead of one?
If you are doing memory intensive stuff, you could be hitting cache capacity.
You could maybe test this with mock algorithm which just processes same small bit if data over and over so it all should fit in cache.
If it indeed is cache, possible solutions could be making the threads work on same data somehow (like different parts of small data window), or just tweaking the algorithm to be more local (like in sorting, merge sort is generally slower than quick sort, but it is more cache friendly which still makes it better in some cases).
Are your threads reading and writing to items close together in memory? Then you're probably running into false sharing. If thread 1 works with data[1] and thread2 works with data[2], then even though in an ideal world we know that two consecutive reads of data[2] by thread2 will always produce the same result, in the actual world, if thread1 updates data[1] sometime between those two reads, then the CPU will mark the cache as dirty and update it. http://msdn.microsoft.com/en-us/magazine/cc872851.aspx. To solve it, make sure the data each thread is working with is adequately far away in memory from the data the other threads are working with.
That could give you a performance boost, but likely won't get you to 16x—there are lots of things going on under the hood and you'll just have to knock them out one-by-one. And really it's not that your algorithm is running at 30% speed when multithreaded; it's more that your single-threaded algorithm is running at 300% speed, enabled by all sorts of CPU and caching awesomeness that running multithreaded has a harder time taking advantage of. So there's nothing to be "embarrassed" about. But with some diligence, you can perhaps get the multithreaded version working at nearly 300% speed as well.
Also, if you're counting hyperthreaded cores as real cores, well, they're not. They only allow threads to swap really fast when one is blocked. But they'll never let you run at double speed unless your threads are getting blocked half the time anyway, in which case that already means you have opportunity for speedup.
Simple question.
How do you decide what the optimal value for MaxDegreeOfParallelism is for any given algorithm? What are the factors to consider and what are the trade-offs?
I think it depends, if all your tasks are "CPU bound" it would be probably equal to the number of CPUs in your machine. Nevertheless if you tasks are "IO bound" you can to start to increase the number.
When the CPU has to switch from one thread to other (context switch) it has a cost, so if you use too much threads and the CPU is switching all the time, you decrease the performance. In the other hand, if you limit that parameter too much, and the operations are long "IO bound" operations, and the CPU is idle a lot of time waiting for those tasks to complete... you are not doing the most with your machine's resources (and that is what multithreading is about)
I think it depends on each algorithm as #Amdahls's Law has pointed out, and there is not a master rule of thumb you can follow, you will have to plan it and tun it :D
Cheers.
For local compute intensive processes you should try two values;
number of physical cores or processors
number of virtual cores (physical including hyperthreading)
One of these is optimal in my experience. Sometimes using hyperthreading slows down, usually it helps. In C# use Environment.ProcessorCount to find the number of cores including hyperthreading fake cores. Actual physical cores is more difficult to determine. Check other questions for this.
For processes that have to wait for resources (db queries, file retrieval) scaling up can help. If you have to wait 80% of the time a process is busy the rule of thumb is to start 5 threads for it so one thread will be busy at any time. Maximizing the 5x20% each process requires locally.
On a single processor, Will multi-threading increse the speed of the calculation. As we all know that, multi-threading is used for Increasing the User responsiveness and achieved by sepating UI thread and calculation thread. But lets talk about only console application. Will multi-threading increases the speed of the calculation. Do we get culculation result faster when we calculate through multi-threading.
what about on multi cores, will multi threading increse the speed or not.
Please help me. If you have any material to learn more about threading. please post.
Edit:
I have been asked a question, At any given time, only one thread is allowed to run on a single core. If so, why people use multithreading in a console application.
Thanks in advance,
Harsha
In general terms, no it won't speed up anything.
Presumably the same work overall is being done, but now there is the overhead of additional threads and context switches.
On a single processor with HyperThreading (two virtual processors) then the answer becomes "maybe".
Finally, even though there is only one CPU perhaps some of the threads can be pushed to the GPU or other hardware? This is kinda getting away from the "single processor" scenario but could technically be way of achieving a speed increase from multithreading on a single core PC.
Edit: your question now mentions multithreaded apps on a multicore machine.
Again, in very general terms, this will provide an overall speed increase to your calculation.
However, the increase (or lack thereof) will depend on how parallelizable the algorithm is, the contention for memory and cache, and the skill of the programmer when it comes to writing parallel code without locking or starvation issues.
Few threads on 1 CPU:
may increase performance in case you continue with another thread instead of waiting for I/O bound operation
may decrease performance if let say there are too many threads and work is wasted on context switching
Few threads on N CPUs:
may increase performance if you are able to cut job in independent chunks and process them in independent manner
may decrease performance if you rely heavily on communication between threads and bus becomes a bottleneck.
So actually it's very task specific - you can parallel one things very easy while it's almost impossible for others. Perhaps it's a bit advanced reading for new person but there are 2 great resources on this topic in C# world:
Joe Duffy's web log
PFX team blog - they have a very good set of articles for parallel programming in .NET world including patterns and practices.
What is your calculation doing? You won't be able to speed it up by using multithreading if it a processor bound, but if for some reason your calculation writes to disk or waits for some other sort of IO you may be able to improve performance using threading. However, when you say "calculation" I assume you mean some sort of processor intensive algorithm, so adding threads is unlikely to help, and could even slow you down as the context switch between threads adds extra work.
If the task is compute bound, threading will not make it faster unless the calculation can be split in multiple independent parts. Even so you will only be able to achieve any performance gains if you have multiple cores available. From the background in your question it will just add overhead.
However, you may still want to run any complex and long running calculations on a separate thread in order to keep the application responsive.
No, no and no.
Unless you write parallelizing code to take advantage of multicores, it will always be slower if you have no other blocking functions.
Exactly like the user input example, one thread might be waiting for a disk operation to complete, and other threads can take that CPU time.
As described in the other answers, multi-threading on a single core won't give you any extra performance (hyperthreading notwithstanding). However, if your machine sports an Nvidia GPU you should be able to use the CUDA to push calculations to the GPU. See http://www.hoopoe-cloud.com/Solutions/CUDA.NET/Default.aspx and C#: Perform Operations on GPU, not CPU (Calculate Pi).
Above mention most.
Running multiple threads on one processor can increase performance, if you can manage to get more work done at the same time, instead of let the processor wait between different operations. However, it could also be a severe loss of performance due to for example synchronization or that the processor is overloaded and cant step up to the requirements.
As for multiple cores, threading can improve the performance significantly. However, much depends on finding the hotspots and not overdo it. Using threads everywhere and the need of synchronization can even lower the performance. Optimizing using threads with multiple cores takes a lot of pre-studies and planning to get a good result. You need for example to think about how many threads to be use in different situations. You do not want the threads to sit and wait for information used by another thread.
http://www.intel.com/intelpress/samples/mcp_samplech01.pdf
https://computing.llnl.gov/tutorials/parallel_comp/
https://computing.llnl.gov/tutorials/pthreads/
http://en.wikipedia.org/wiki/Superscalar
http://en.wikipedia.org/wiki/Simultaneous_multithreading
I have been doing some intensive C++ mathematical simulation runs using 24 core servers. If I run 24 separate simulations in parallel on the 24 cores of a single server, then I get a runtime for each of my simulations of say X seconds.
The bizarre thing I have noticed is that, when running only 12 simulations, using 12 of the 24 cores, with the other 12 cores idle, then each of the simulations runs at a runtime of Y seconds, where Y is much greater than X! When viewing the task manager graph of the processor usage, it is obvious that a process does not stick to only one core, but alternates between a number of cores. That is to say, the switching between cores to use all the cores slows down the calculation process.
The way I maintained the runtime when running only 12 simulations, is to run another 12 "junk" simulations on the side, using the remaining 12 cores!
Conclusion: When using multi-cores, use them all at 100%, for lower utilisation, the runtime increases!
For single core CPU,
Actually the performance depends on the job you are referring.
In your case, for calculation done by CPU, in that case OverClocking would help if your parentBoard supports it. Otherwise there is no way for CPU to do calculations that are faster than the speed of CPU.
For the sake of Multicore CPU
As the above answers say, if properly designed the performance may increase, if all cores are fully used.
In single core CPU, if the threads are implemented in User Level then multithreading wont matter if there are blocking system calls in the thread, like an I/O operation. Because kernel won't know about the userlevel threads.
So if the process does I/O then you can implement the threads in Kernel space and then you can implement different threads for different job.
(The answer here is on theory based.)
Even a CPU bound task might run faster multi-threaded if properly designed to take advantage of cache memory and pipelineing done by the processor. Modern processors spend a lot of time
twiddling their thumbs, even when nominally fully "busy".
Imagine a process that used a small chunk of memory very intensively. Processing
the same chunk of memory 1000 times would be much faster than processing 1000 chunks
of similar memory.
You could certainly design a multi threaded program that would be faster than a single thread.
Treads don't increase performance. Threads sacrifice performance in favor of keeping parts of the code responsive.
The only exception is if you are doing a computation that is so parallelizeable that you can run different threads on different cores (which is the exception, not the rule).
recently I record several hours of .NET Memory counters of a WCF service. The service is hosted in IIS on a Win2k8, 8 core, x64 with 20GB ram.
I can see the GC being pretty healthy, performing a full collection only approx. every 2hrs!
I noticed that in the very same time period, the number of physical and logical threads increases. When the full collection occurs, the number of physical and logical threads drops back and continues to raise again to the same level.
Why does the GC full collection cycle and the drop in threads correlate?
Why is the number of threads continuously increasing?
This is pure ASP.NET/WCF threading model. No custom threads being spawned etc.
Thanks,
Alex
It looks like the answer was indeed pretty simple: Threads get created and die as the thread pool is grown and shrinked over time. Normally, those threads have a relatively long life-time which means they end up in generation 2. This is why they naturally only cleaned up on a full GC cycle.
The interesting bit is that a very well memory managed application can end up with loads of dead threads which I was able to observe in WinDbg. This is because a Full GC might only happen after several hours or even days! Some MS blog said that those dead threads are not a really nice thing to have as the corresponding unmanaged C++ thread object seems to be quite resource intensive (cannot find the damn blog anymore unfortunatly....). I am wondering if this is one of the rare cases where an induced Full GC would make sense in a certain time interval.