How can I get faster processing speeds for a single thread by combining multiple CPU cores, like training a custom neural network (not tensorflow) on a Google Compute Engine n1-highmem-64 machine type that has 64 CPU cores? Cluster computers or what? Not sure where to start... thanks!
Well you are asking for faster speeds on a single thread, but with multiple cores.
The only feasible way to get faster processing speeds off of a single thread, which is owned by a single core, is overclocking. You could also get a better chipset by getting newer cores.
It would be infeasible to accomplish this straightforward, you would likely have to patch the firmware to several of your components allowing them to communicate across cpus on a single thread utilizing an L3 cache or something.... very infeasible.
The opposite of that would be the way to go.
Multi-threading is used for processing different pieces of data concurrently on multiple cores.
General Purpose GPU use is for performing the same operation on a large set of data by farming the computation to the GPU. It has increased overhead time, but will give good results when the input is big enough.
It's funny that you mention not TensorFlow, because it actually implements both of these.
Even if you were able to implement something like this, it would probably just thrash over atomic locks unless you threaded it anyway.
Edit
If you are looking to use software as a service, Amazon (https://aws.amazon.com/tensorflow/ and other companies) have a range of services that are compatible with various deep-learning / machine-learning frameworks out of the box.
Related
If a single threaded process is busy and uses 100% of a single core it seems like Windows is switching this process between the cores, because in Task Managers core overview all cores are equal used.
Why does Windows do that? Isn't this destroying L1/L2 caches?
There are advantages to pinning a process to one core, primarily caching which you already mentioned.
There are also disadvantages -- you get unequal heating, which can create mechanical stresses that do not improve the expected lifetime of the silicon die.
To avoid this, OSes tend to keep all cores at equal utilization. When there's only one active thread, it will have to be moved and invalidate caches. As long as this is done infrequently (in CPU time), the impact of the extra cache misses during migration is negligible.
For example, the abstract of "Energy and thermal tradeoffs in hardware-based load balancing for clustered multi-core architectures implementing power gating" explicitly lists this as a design goal of scheduling algorithms (emphasis mine):
In this work, a load-balancing technique for these clustered multi-core architectures is presented that provides both a low overhead in energy and an a smooth temperature distribution across the die, increasing the reliability of the processor by evenly stressing the cores.
Spreading the heat dissipation throughout the die is also essential for techniques such as Turbo Boost, where cores are clocked temporarily at a rate that is unsustainable long term. By moving load to a different core regularly, the average heat dissipation remains sustainable even though the instantaneous power is not.
Your process may be the only one doing a lot of work, but it is not the only thing running. There are lots of other processes that need to run occasionally. When your process gets evicted and eventually re-scheduled, the core on which it was running previously might not be available. It's better to run the waiting process immediately on a free core than to wait for the previous core to be available (and in any case its data will likely have been bumped from the caches by the other thread).
In addition, modern CPUs allow all the cores in a package to share high-level caches. See the "Smart Cache" feature in this Intel Core i5 spec sheet. You still lose the lower-level cache(s) on core switch, but those are small and will probably churn somewhat anyway if you're running more than just a small tight loop.
We've just bought a 32-core Opteron machine, and the speedups we get are a little disappointing: beyond about 24 threads we see no speedup at all (actually gets slower overall) and after about 6 threads it becomes significantly sub-linear.
Our application is very thread-friendly: our job breaks down into about 170,000 little tasks which can each be executed separately, each taking 5-10 seconds. They all read from the same memory-mapped file of size about 4Gb. They make occasional writes to it, but it might be 10,000 reads to each write - we just write a little bit of data at the end of each of the 170,000 tasks. The writes are lock-protected. Profiling shows that the locks are not a problem. The threads use a lot of JVM memory each in non-shared objects and they make very little access to shared JVM objects and of that, only a small percentage of accesses involve writes.
We're programming in Java, on Linux, with NUMA enabled. We have 128Gb RAM. We have 2 Opteron CPU's (model 6274) of 16 cores each. Each CPU has 2 NUMA nodes. The same job running on an Intel quad-core (i.e. 8 cores) scaled nearly linearly up to 8 threads.
We've tried replicating the read-only data to have one-per-thread, in the hope that most lookups can be local to a NUMA node, but we observed no speedup from this.
With 32 threads, 'top' shows the CPU's 74% "us" (user) and about 23% "id" (idle). But there are no sleeps and almost no disk i/o. With 24 threads we get 83% CPU usage. I'm not sure how to interpret 'idle' state - does this mean 'waiting for memory controller'?
We tried turning NUMA on and off (I'm referring to the Linux-level setting that requires a reboot) and saw no difference. When NUMA was enabled, 'numastat' showed only about 5% of 'allocation and access misses' (95% of cache misses were local to the NUMA node). [Edit:] But adding "-XX:+useNUMA" as a java commandline flag gave us a 10% boost.
One theory we have is that we're maxing out the memory controllers, because our application uses a lot of RAM and we think there are a lot of cache misses.
What can we do to either (a) speed up our program to approach linear scalability, or (b) diagnose what's happening?
Also: (c) how do I interpret the 'top' result - does 'idle' mean 'blocked on memory controllers'? and (d) is there any difference in the characteristics of Opteron vs Xeon's?
I also have a 32 core Opteron machine, with 8 NUMA nodes (4x6128 processors, Mangy Cours, not Bulldozer), and I have faced similar issues.
I think the answer to your problem is hinted at by the 2.3% "sys" time shown in top. In my experience, this sys time is the time the system spends in the kernel waiting for a lock. When a thread can't get a lock it then sits idle until it makes its next attempt. Both the sys and idle time are a direct result of lock contention. You say that your profiler is not showing locks to be the problem. My guess is that for some reason the code causing the lock in question is not included in the profile results.
In my case a significant cause of lock contention was not the processing I was actually doing but the work scheduler that was handing out the individual pieces of work to each thread. This code used locks to keep track of which thread was doing which piece of work. My solution to this problem was to rewrite my work scheduler avoiding mutexes, which I have read do not scale well beyond 8-12 cores, and instead use gcc builtin atomics (I program in C on Linux). Atomic operations are effectively a very fine grained lock that scales much better with high core counts. In your case if your work parcels really do take 5-10s each it seems unlikely this will be significant for you.
I also had problems with malloc, which suffers horrible lock issues in high core count situations, but I can't, off the top of my head, remember whether this also led to sys & idle figures in top, or whether it just showed up using Mike Dunlavey's debugger profiling method (How can I profile C++ code running in Linux?). I suspect it did cause sys & idle problems, but I draw the line at digging through all my old notes to find out :) I do know that I now avoid runtime mallocs as much as possible.
My best guess is that some piece of library code you are using implements locks without your knowledge, is not included in your profiling results, and is not scaling well to high core-count situations. Beware memory allocators!
I'm sure the answer will lie in a consideration of the hardware architecture. You have to think of multi core computers as if they were individual machines connected by a network. In fact that's all that Hypertransport and QPI are.
I find that to solve these scalability problems you have to stop thinking in terms of shared memory and start adopting the philosophy of Communicating Sequential Processes. It means thinking very differently, ie imagine how you would write the software if your hardware was 32 single core machines connected by a network. Modern (and ancient) CPU architectures are not designed to give unfettered scaling of the sort you're after. They are designed to allow many different processes to get on with processing their own data.
Like everything else in computing these things go in fashions. CSP dates back to the 1970s, but the very modern and Java derived Scala is a popular embodiment of the concept. See this section on Scala concurrency on Wikipedia.
What the philosophy of CSP does is force you to design a data distribution scheme that fits your data and the problem you're solving. That's not necessarily easy, but if you manage it then you have a solution that will scale very well indeed. Scala may make it easier to develop.
Personally I do everything in CSP and in C. It's allowed me to develop a signal processing application that scales perfectly linearly from 8 cores to several thousand cores (the limit being how big my room is).
The first thing you're going to have to do is actually use NUMA. It isn't a magic setting that you turn on, you have to exploit it in your software's architecture. I don't know about Java, but in C one would bind a memory allocation to a specific core's memory controller (aka memory affinity), and similarly for threads (core affinity) in cases where the OS doesn't get the hint.
I presume that your data doesn't break down into 32 neat, discrete chunks? It's difficult to give advice without knowing exactly the data flows implicit in your program. But think about it in terms of data flow. Draw it out even; Data Flow Diagrams are useful for this (another ancient graphical formal notation). If your picture shows all your data going through a single object (eg through a single memory buffer) then it's going to be slow...
I assume you have optimized your locks, and synchronization made a minimum. In such a case, it still depends a lot on what libraries you are using to program in parallel.
One issue that can happen even if you have no synchronization issue, is memory bus congestion. This is very nasty and difficult to get rid of.
All I can suggest is somehow make your tasks bigger and create fewer tasks. This depends highly on the nature of your problem. Ideally you want as many tasks as the number of cores/threads, but this is not easy (if possible) to achieve.
Something else that can help is to give more heap to your JVM. This will reduce the need to run Garbage Collector frequently, and speeds up a little.
does 'idle' mean 'blocked on memory controllers'
No. You don't see that in top. I mean if the CPU is waiting for memory access, it will be shown as busy. If you have idle periods, it is either waiting for a lock, or for IO.
I'm the Original Poster. We think we've diagnosed the issue, and it's not locks, not system calls, not memory bus congestion; we think it's level 2/3 CPU cache contention.
To reiterate, our task is embarrassingly parallel so it should scale well. However, one thread has a large amount of CPU cache it can access, but as we add more threads, the amount of CPU cache each process can access gets lower and lower (the same amount of cache divided by more processes). Some levels on some architectures are shared between cores on a die, some are even shared between dies (I think), and it may help to get "down in the weeds" with the specific machine you're using, and optimise your algorithms, but our conclusion is that there's not a lot we can do to achieve the scalability we thought we'd get.
We identified this as the cause by using 2 different algorithms. The one which accesses more level 2/3 cache scales much worse than the one which does more processing with less data. They both make frequent accesses to the main data in main memory.
If you haven't tried that yet: Look at hardware-level profilers like Oracle Studio has (for CentOS, Redhat, and Oracle Linux) or if you are stuck with Windows: Intel VTune. Then start looking at operations with suspiciously high clocks per instruction metrics. Suspiciously high mean a lot higher than the same code on a single-numa, single-L3-cache machine (like current Intel desktop CPUs).
Will the current trend of adding cores to computers continue? Or is there some theoretical or practical limit to the number of cores that can be served by one set of memory?
Put another way: is the high powered desktop computer of the future apt to have 1024 cores using one set of memory, or is it apt to have 32 sets of memory, each accessed by 32 cores?
Or still another way: I have a multi-threaded program that runs well on a 4-core machine, using a significant amount of the total CPU. As this program grows in size and does more work, can I be reasonably confident more powerful machines will be available to run it? Or should I be thinking seriously about running multiple sessions on multiple machines (or at any rate multiple sets of memory) to get the work done?
In other words, is a purely multithreaded approach to design going to leave me in a dead end? (As using a single threaded approach and depending on continued improvements in CPU speed years back would have done?) The program is unlikely to be run on a machine costing more than, say $3,000. If that machine cannot do the work, the work won't get done. But if that $3,000 machine is actually a network of 32 independent computers (though they may share the same cooling fan) and I've continued my massively multithreaded approach, the machine will be able to do the work, but the program won't, and I'm going to be in an awkward spot.
Distributed processing looks like a bigger pain than multithreading was, but if that might be in my future, I'd like some warning.
Will the current trend of adding cores to computers continue?
Yes, the GHz race is over. It's not practical to ramp the speed any more on the current technology. Physics has gotten in the way. There may be a dramatic shift in the technology of fabricating chips that allows us to get round this, but it's not obviously 'just around the corner'.
If we can't have faster cores, the only way to get more power is to have more cores.
Or is there some theoretical or practical limit to the number of cores that can be served by one set of memory?
Absolutely there's a limit. In a shared memory system the memory is a shared resource and has a limited amount of bandwidth.
Max processes = (Memory Bandwidth) / (Bandwidth required per process)
Now - that 'Bandwidth per process' figure will be reduced by caches, but caches become less efficient if they have to be coherent with one another because everyone is accessing the same area of memory. (You can't cache a memory write if another CPU may need what you've written)
When you start talking about huge systems, shared resources like this become the main problem. It might be memory bandwidth, CPU cycles, hard drive access, network bandwidth. It comes down to how the system as a whole is structured.
You seem to be really asking for a vision of the future so you can prepare. Here's my take.
I think we're going to see a change in the way software developers see parallelism in their programs. At the moment, I would say that a lot of software developers see the only way of using multiple threads is to have lots of them doing the same thing. The trouble is they're all contesting for the same resources. This then means lots of locking needs to be introduced, which causes performance issues, and subtle bugs which are infuriating and time consuming to solve.
This isn't sustainable.
Manufacturing worked out at the beginning of the 20th Century, the fastest way to build lots of cars wasn't to have lots of people working on one car, and then, when that one's done, move them all on to the next car. It was to split the process of building the car down into lots of small jobs, with the output of one job feeding the next. They called it assembly lines. In hardware design it's called pipe-lining, and I'll think we'll see software designs move to it more and more, as it minimizes the problem of shared resources.
Sure - There's still a shared resource on the output of one stage and the input of the next, but this is only between two threads/processes and is much easier to handle. Standard methods can also be adopted on how these interfaces are made, and message queueing libraries seem to be making big strides here.
There's not one solution for all problems though. This type of pipe-line works great for high throughput applications that can absorb some latency. If you can't live with the latency, you have no option but to go the 'many workers on a single task' route. Those are the ones you ideally want to be throwing at SIMD machines/Array processors like GPUs, but it only really excels with a certain type of problem. Those problems are ones where there's lots of data to process in the same way, and there's very little or no dependency between data items.
Having a good grasp of message queuing techniques and similar for pipelined systems, and utilising fine grained parallelism on GPUs through libraries such as OpenCL, will give you insight at both ends of the spectrum.
Update: Multi-threaded code may run on clustered machines, so this issue may not be as critical as I thought.
I was carefully checking out the Java Memory Model in the JLS, chapter 17, and found it does not mirror the typical register-cache-main memory model of most computers. There were opportunities there for a multi-memory machine to cleanly shift data from one memory to another (and from one thread running on one machine to another running on a different one). So I started searching for JVMs that would run across multiple machines. I found several old references--the idea has been out there, but not followed through. However, one company, Terracotta, seems to have something, if I'm reading their PR right.
At any rate, it rather seems that when PC's typically contain several clustered machines, there's likely to be a multi-machine JVM for them.
I could find nothing outside the Java world, but Microsoft's CLR ought to provide the same opportunities. C and C++ and all the other .exe languages might be more difficult. However, Terracotta's websites talk more about linking JVM's rather than one JVM on multiple machines, so their tricks might work for executable langauges also (and maybe the CLR, if needed).
I think the topic says it all. What's the difference, if any, between parallel and multicore programming? Thanks.
Mutli-core is a kind of parallel programming. In particular, it is a kind of MIMD setup where the processing units aren't distributed, but rather share a common memory area, and can even share data like a MISD setup if need be. I believe it is even disctinct from multi-processing, in that a multi-core setup can share some level of caches, and thus cooperate more efficiently than CPUs on different cores.
General parallel programing would also include SIMD systems (like your GPU), and distributed systems.
The difference isn't in approach, just in the hardware the software runs on. Parallel programming is taking a problem and spliting the workload into smaller pieces that can be processed in parallel(Divide and Conquer type problems, etc.) or functions that can run independently of each other. Place that software on a multi-core piece of hardware and it will be optimized by the OS to run on the different cores. This gives it a better performance because each thread you create to do concurrent work can now run without consuming CPU cycles on a single processor/core.
Multicore systems are a subset of parallel systems. Different systems will have different memory architectures, each with their own set of challenges. How does one system deal with cache coherency? Is NUMA involved, etc. etc.
With the recent buzz on multicore programming is anyone exploring the possibilities of using MPI ?
I've used MPI extensively on large clusters with multi-core nodes. I'm not sure if it's the right thing for a single multi-core box, but if you anticipate that your code may one day scale larger than a single chip, you might consider implementing it in MPI. Right now, nothing scales larger than MPI. I'm not sure where the posters who mention unacceptable overheads are coming from, but I've tried to give an overview of the relevant tradeoffs below. Read on for more.
MPI is the de-facto standard for large-scale scientific computation and it's in wide use on multicore machines already. It is very fast. Take a look at the most recent Top 500 list. The top machines on that list have, in some cases, hundreds of thousands of processors, with multi-socket dual- and quad-core nodes. Many of these machines have very fast custom networks (Torus, Mesh, Tree, etc) and optimized MPI implementations that are aware of the hardware.
If you want to use MPI with a single-chip multi-core machine, it will work fine. In fact, recent versions of Mac OS X come with OpenMPI pre-installed, and you can download an install OpenMPI pretty painlessly on an ordinary multi-core Linux machine. OpenMPI is in use at Los Alamos on most of their systems. Livermore uses mvapich on their Linux clusters. What you should keep in mind before diving in is that MPI was designed for solving large-scale scientific problems on distributed-memory systems. The multi-core boxes you are dealing with probably have shared memory.
OpenMPI and other implementations use shared memory for local message passing by default, so you don't have to worry about network overhead when you're passing messages to local processes. It's pretty transparent, and I'm not sure where other posters are getting their concerns about high overhead. The caveat is that MPI is not the easiest thing you could use to get parallelism on a single multi-core box. In MPI, all the message passing is explicit. It has been called the "assembly language" of parallel programming for this reason. Explicit communication between processes isn't easy if you're not an experienced HPC person, and there are other paradigms more suited for shared memory (UPC, OpenMP, and nice languages like Erlang to name a few) that you might try first.
My advice is to go with MPI if you anticipate writing a parallel application that may need more than a single machine to solve. You'll be able to test and run fine with a regular multi-core box, and migrating to a cluster will be pretty painless once you get it working there. If you are writing an application that will only ever need a single machine, try something else. There are easier ways to exploit that kind of parallelism.
Finally, if you are feeling really adventurous, try MPI in conjunction with threads, OpenMP, or some other local shared-memory paradigm. You can use MPI for the distributed message passing and something else for on-node parallelism. This is where big machines are going; future machines with hundreds of thousands of processors or more are expected to have MPI implementations that scale to all nodes but not all cores, and HPC people will be forced to build hybrid applications. This isn't for the faint of heart, and there's a lot of work to be done before there's an accepted paradigm in this space.
I would have to agree with tgamblin. You'll probably have to roll your sleeves up and really dig into the code to use MPI, explicitly handling the organization of the message-passing yourself. If this is the sort of thing you like or don't mind doing, I would expect that MPI would work just as well on multicore machines as it would on a distributed cluster.
Speaking from personal experience... I coded up some C code in graduate school to do some large scale modeling of electrophysiologic models on a cluster where each node was itself a multicore machine. Therefore, there were a couple of different parallel methods I thought of to tackle the problem.
1) I could use MPI alone, treating every processor as it's own "node" even though some of them are grouped together on the same machine.
2) I could use MPI to handle data moving between multicore nodes, and then use threading (POSIX threads) within each multicore machine, where processors share memory.
For the specific mathematical problem I was working on, I tested two formulations first on a single multicore machine: one using MPI and one using POSIX threads. As it turned out, the MPI implementation was much more efficient, giving a speed-up of close to 2 for a dual-core machine as opposed to 1.3-1.4 for the threaded implementation. For the MPI code, I was able to organize operations so that processors were rarely idle, staying busy while messages were passed between them and masking much of the delay from transferring data. With the threaded code, I ended up with a lot of mutex bottlenecks that forced threads to often sit and wait while other threads completed their computations. Keeping the computational load balanced between threads didn't seem to help this fact.
This may have been specific to just the models I was working on, and the effectiveness of threading vs. MPI would likely vary greatly for other types of parallel problems. Nevertheless, I would disagree that MPI has an unwieldy overhead.
No, in my opinion it is unsuitable for most processing you would do on a multicore system. The overhead is too high, the objects you pass around must be deeply cloned, and passing large objects graphs around to then run a very small computation is very inefficient. It is really meant for sharing data between separate processes, most often running in separate memory spaces, and most often running long computations.
A multicore processor is a shared memory machine, so there are much more efficient ways to do parallel processing, that do not involve copying objects and where most of the threads run for a very small time. For example, think of a multithreaded Quicksort. The overhead of allocating memory and copying the data to a thread before it can be partioned will be much slower with MPI and an unlimited number of processors than Quicksort running on a single processor.
As an example, in Java, I would use a BlockingQueue (a shared memory construct), to pass object references between threads, with very little overhead.
Not that it does not have its place, see for example the Google search cluster that uses message passing. But it's probably not the problem you are trying to solve.
MPI is not inefficient. You need to break the problem down into chunks and pass the chunks around and reorganize when the result is finished per chunk. No one in the right mind would pass around the whole object via MPI when only a portion of the problem is being worked on per thread. Its not the inefficiency of the interface or design pattern thats the inefficiency of the programmers knowledge of how to break up a problem.
When you use a locking mechanism the overhead on the mutex does not scale well. this is due to the fact that the underlining runqueue does not know when you are going to lock the thread next. You will perform more kernel level thrashing using mutex's than a message passing design pattern.
MPI has a very large amount of overhead, primarily to handle inter-process communication and heterogeneous systems. I've used it in cases where a small amount of data is being passed around, and where the ratio of computation to data is large.
This is not the typical usage scenario for most consumer or business tasks, and in any case, as a previous reply mentioned, on a shared memory architecture like a multicore machine, there are vastly faster ways to handle it, such as memory pointers.
If you had some sort of problem with the properties describe above, and you want to be able to spread the job around to other machines, which must be on the same highspeed network as yourself, then maybe MPI could make sense. I have a hard time imagining such a scenario though.
I personally have taken up Erlang( and i like to so far). The messages based approach seem to fit most of the problem and i think that is going to be one of the key item for multi core programming. I never knew about the overhead of MPI and thanks for pointing it out
You have to decide if you want low level threading or high level threading. If you want low level then use pThread. You have to be careful that you don't introduce race conditions and make threading performance work against you.
I have used some OSS packages for (C and C++) that are scalable and optimize the task scheduling. TBB (threading building blocks) and Cilk Plus are good and easy to code and get applications of the ground. I also believe they are flexible enough integrate other thread technologies into it at a later point if needed (OpenMP etc.)
www.threadingbuildingblocks.org
www.cilkplus.org