How to profile OpenMP bottlenecks - c++

I have a loop that has been parallelized by OpenMP, but due to the nature of the task, there are 4 critical clauses.
What would be the best way to profile the speed up and find out which of the critical clauses (or maybe non-critical(!) ) take up the most time inside the loop?
I use Ubuntu 10.04 with g++ 4.4.3

Scalasca is a nice tool for profiling OpenMP (and MPI) codes and analyzing the results. Tau is also very nice but much harder to use. The intel tools, like the vtune, are also good but very expensive.

Arm MAP has OpenMP and pthreads profiling - and works without needing to instrument or modify your source code. You can see synchronization issues and where threads are spending time to the source line level. The OpenMP profiling blog entry is worth reading.
MAP is widely used for high performance computing as it is also profiles multiprocess applications such as MPI.

OpenMP includes the functions omp_get_wtime() and omp_get_wtick() for measuring timing performance (docs here), I would recommend using these.
Otherwise try a profiler. I prefer the google CPU profiler which can be found here.
There is also the manual way described in this answer.

There is also the ompP tool which I have used a number of times in the last ten years. I have found it to be really useful to identify and quantify load imbalance and parallel/serial regions. The web page seems to be down now but I also found it on web archive earlier this year.
edit: updated home directory

Related

Can bad vectorized code impact on scalability?

I have parallelized an already existing code for computer vision applications using OpenMP. I think that I well designed it because:
The workload is well-balanced
There is no synchronization/locking mechanism
I parallelized the outer most loops
All the cores are used for most of the time (there are no idle cores)
There is enough work for each thread
Now, the application doesn't scale when using many cores, e.g. it doesn't scale well after 15 cores.
The code uses external libraries (i.e. OpenCV and IPP) where the code is already optimized and vectorized, while I manually vectorized some portions of the code as best as I could. However, according to Intel Advisor, the code isn't well vectorized, but there is no much left to do: I already vectorized the code where I could and I can't improve the external libraries.
So my question is: is it possible that vectorization is the reason why the code doesn't scale well at some point? If so, why?
In line with comments from Adam Nevraumont, VTune Amplifier can do a lot to pinpoint memory bandwidth issues: https://software.intel.com/en-us/vtune-amplifier-help-memory-access-analysis.
It may be useful to start at a higher level of analysis than that though, like looking at hot spots. If it turns out that most of your time is spent in OpenCV or similar like you're concerned about, finding that out early might save some time vs. digging into memory bottlenecks directly.

How to profile lock contentions under g++/std::mutex?

Question
Are there any open-source tools or does anyone have any techniques/code for profiling the degree of std::mutex contentions in running code?
I would like to count the percentage of lock contention at the granularity (either by time or number) of each std::mutex instance. If there is a drop-in tool that doesn't require recoding, that would be even better.
I am looking for a technique that will work with std::thread and g++ : at the exit of the application, I would like to dump out a profile of mutex contention statistics into a log file, so that I can monitor the quality of threading code under actual running contexts.
Note
I have seen this thread. Unfortunately, the answers either require a pile of cash or run on Windows.
I recommend something like AMD CodeXL or Intel VTune. CodeXL is free; Intel VTune has free academic license if that's applicable to you, or you can try a 30-day trial. Both of them work in Linux.
At the most basic level, these tools can identify hotspots by eg, measuring how much time you are spending inside methods of std::mutex. There are other more advanced analysis techniques/tools included in each tool that may help you even further. You don't need to change your code at all, although you may need to check that you compiled with debug symbols and/or haven't stripped the binaries. You will also probably want to stay away from extreme optimization levels like -O3, and stick to -O1, -O2 or -Og.
PS: As will all optimization inquiries, I must remind you to always measure where your performance problems actually are before you start optimizing. No matter how worried you are about lock contention, validate your concerns with a profiler before you spend huge efforts trying to alleviate whatever lock contention you may or may not be having.

Testing performance of C++ code

What free tools could I use to test the performance of C++ code in Linux? Basically I want to identify the bottleneck of the code and improve on the performance. My application mainly involves computational code using the data from the network. So I would like to improve the speed of execution of the code.
Thanks.
For typical performance benchmarking this is what i use.
gprof/oprofile - for CPU intensive profiling of your code.
netstat/ethereal - for network statistics
iostat/sar - for I/O
vmstat - for memory
mpstat/sar - for cpu usage
Now u can isolate the problems based on the output of these tools.
For eg:- if I/O is constant and within limits u can eliminate I/O as a problem.
If CPU usage is heavy as shown my mpstat then get into profiling using gprof/oprofile.
Without the use of all of them together for different runs, its difficult to identify the bottleneck.
Note: U can write a script to run all of them together and store the results in designated folders for each run.
I recommend valgrind for
cpu usage, callgrind submodule (source line granularity)
memory leaks
building call graphs
some advanced issues like finding problems in multithread locking mechanism
The callgrind output can be visually displayed via KCacheGrind.
By far the best profiler for Linux that I know of is Zoom. Although it's a commercial product it's not too expensive and you can get a free 30 day evaluation licence on request
As #Paul says, give Zoom a try.
Personally, I use this method, which works for these reasons, and Zoom approximates it. It is a technique that some programmers have independently discovered.
I'm also told OProfile can do it, but you have to know what you need to make it do.

How to profile multi-threaded C++ application on Linux?

I used to do all my Linux profiling with gprof.
However, with my multi-threaded application, it's output appears to be inconsistent.
Now, I dug this up:
http://sam.zoy.org/writings/programming/gprof.html
However, it's from a long time ago and in my gprof output, it appears my gprof is listing functions used by non-main threads.
So, my questions are:
In 2010, can I easily use gprof to profile multi-threaded Linux C++ applications? (Ubuntu 9.10)
What other tools should I look into for profiling?
Edit: added another answer on poor man's profiler, which IMHO is better for multithreaded apps.
Have a look at oprofile. The profiling overhead of this tool is negligible and it supports multithreaded applications---as long as you don't want to profile mutex contention (which is a very important part of profiling multithreaded applications)
Have a look at poor man's profiler. Surprisingly there are few other tools that for multithreaded applications do both CPU profiling and mutex contention profiling, and PMP does both, while not even requiring to install anything (as long as you have gdb).
Try modern linux profiling tool, the perf (perf_events): https://perf.wiki.kernel.org/index.php/Tutorial and http://www.brendangregg.com/perf.html:
perf record ./application
# generates profile file perf.data
perf report
Have a look at Valgrind.
A Paul R said, have a look at Zoom. You can also use lsstack, which is a low-tech approach but surprisingly effective, compared to gprof.
Added: Since you clarified that you are running OpenGL at 33ms, my prior recommendation stands. In addition, what I personally have done in situations like that is both effective and non-intuitive. Just get it running with a typical or problematic workload, and just stop it, manually, in its tracks, and see what it's doing and why. Do this several times.
Now, if it only occasionally misbehaves, you would like to stop it only while it's misbehaving. That's not easy, but I've used an alarm-clock interrupt set for just the right delay. For example, if one frame out of 100 takes more than 33ms, at the start of a frame, set the timer for 35ms, and at the end of a frame, turn it off. That way, it will interrupt only when the code is taking too long, and it will show you why. Of course, one sample might miss the guilty code, but 20 samples won't miss it.
I tried valgrind and gprof. It is a crying shame that none of them work well with multi-threaded applications. Later, I found Intel VTune Amplifier. The good thing is, it handles multi-threading well, works with most of the major languages, works on Windows and Linux, and has many great profiling features. Moreover, the application itself is free. However, it only works with Intel processors.
You can randomly run pstack to find out the stack at a given point. E.g. 10 or 20 times.
The most typical stack is where the application spends most of the time (according to experience, we can assume a Pareto distribution).
You can combine that knowledge with strace or truss (Solaris) to trace system calls, and pmap for the memory print.
If the application runs on a dedicated system, you have also sar to measure cpu, memory, i/o, etc. to profile the overall system.
Since you didn't mention non-commercial, may I suggest Intel's VTune. It's not free but the level of detail is very impressive (and the overhead is negligible).
Putting a slightly different twist on matters, you can actually get a pretty good idea as to what's going on in a multithreaded application using ftrace and kernelshark. Collecting the right trace and pressing the right buttons and you can see the scheduling of individual threads.
Depending on your distro's kernel you may have to build a kernel with the right configuration (but I think that a lot of them have it built in these days).
Microprofile is another possible answer to this. It requires hand-instrumentation of the code, but it seems like it handles multi-threaded code pretty well. And it also has special hooks for profiling graphics pipelines, including what's going on inside the card itself.

How are you taking advantage of Multicore?

As someone in the world of HPC who came from the world of enterprise web development, I'm always curious to see how developers back in the "real world" are taking advantage of parallel computing. This is much more relevant now that all chips are going multicore, and it'll be even more relevant when there are thousands of cores on a chip instead of just a few.
My questions are:
How does this affect your software roadmap?
I'm particularly interested in real stories about how multicore is affecting different software domains, so specify what kind of development you do in your answer (e.g. server side, client-side apps, scientific computing, etc).
What are you doing with your existing code to take advantage of multicore machines, and what challenges have you faced? Are you using OpenMP, Erlang, Haskell, CUDA, TBB, UPC or something else?
What do you plan to do as concurrency levels continue to increase, and how will you deal with hundreds or thousands of cores?
If your domain doesn't easily benefit from parallel computation, then explaining why is interesting, too.
Finally, I've framed this as a multicore question, but feel free to talk about other types of parallel computing. If you're porting part of your app to use MapReduce, or if MPI on large clusters is the paradigm for you, then definitely mention that, too.
Update: If you do answer #5, mention whether you think things will change if there get to be more cores (100, 1000, etc) than you can feed with available memory bandwidth (seeing as how bandwidth is getting smaller and smaller per core). Can you still use the remaining cores for your application?
My research work includes work on compilers and on spam filtering. I also do a lot of 'personal productivity' Unix stuff. Plus I write and use software to administer classes that I teach, which includes grading, testing student code, tracking grades, and myriad other trivia.
Multicore affects me not at all except as a research problem for compilers to support other applications. But those problems lie primarily in the run-time system, not the compiler.
At great trouble and expense, Dave Wortman showed around 1990 that you could parallelize a compiler to keep four processors busy. Nobody I know has ever repeated the experiment. Most compilers are fast enough to run single-threaded. And it's much easier to run your sequential compiler on several different source files in parallel than it is to make your compiler itself parallel. For spam filtering, learning is an inherently sequential process. And even an older machine can learn hundreds of messages a second, so even a large corpus can be learned in under a minute. Again, training is fast enough.
The only significant way I have of exploiting parallel machines is using parallel make. It is a great boon, and big builds are easy to parallelize. Make does almost all the work automatically. The only other thing I can remember is using parallelism to time long-running student code by farming it out to a bunch of lab machines, which I could do in good conscience because I was only clobbering a single core per machine, so using only 1/4 of CPU resources. Oh, and I wrote a Lua script that will use all 4 cores when ripping MP3 files with lame. That script was a lot of work to get right.
I will ignore tens, hundreds, and thousands of cores. The first time I was told "parallel machines are coming; you must get ready" was 1984. It was true then and is true today that parallel programming is a domain for highly skilled specialists. The only thing that has changed is that today manufacturers are forcing us to pay for parallel hardware whether we want it or not. But just because the hardware is paid for doesn't mean it's free to use. The programming models are awful, and making the thread/mutex model work, let alone perform well, is an expensive job even if the hardware is free. I expect most programmers to ignore parallelism and quietly get on about their business. When a skilled specialist comes along with a parallel make or a great computer game, I will quietly applaud and make use of their efforts. If I want performance for my own apps I will concentrate on reducing memory allocations and ignore parallelism.
Parallelism is really hard. Most domains are hard to parallelize. A widely reusable exception like parallel make is cause for much rejoicing.
Summary (which I heard from a keynote speaker who works for a leading CPU manufacturer): the industry backed into multicore because they couldn't keep making machines run faster and hotter and they didn't know what to do with the extra transistors. Now they're desperate to find a way to make multicore profitable because if they don't have profits, they can't build the next generation of fab lines. The gravy train is over, and we might actually have to start paying attention to software costs.
Many people who are serious about parallelism are ignoring these toy 4-core or even 32-core machines in favor of GPUs with 128 processors or more. My guess is that the real action is going to be there.
For web applications it's very, very easy: ignore it. Unless you've got some code that really begs to be done in parallel you can simply write old-style single-threaded code and be happy.
You usually have a lot more requests to handle at any given moment than you have cores. And since each one is handled in its own Thread (or even process, depending on your technology) this is already working in parallel.
The only place you need to be careful is when accessing some kind of global state that requires synchronization. Keep that to a minimum to avoid introducing artificial bottlenecks to an otherwise (almost) perfectly scalable world.
So for me multi-core basically boils down to these items:
My servers have less "CPUs" while each one sports more cores (not much of a difference to me)
The same number of CPUs can substain a larger amount of concurrent users
When the seems to be performance bottleneck that's not the result of the CPU being 100% loaded, then that's an indication that I'm doing some bad synchronization somewhere.
At the moment - doesn't affect it that much, to be honest. I'm more in 'preparation stage', learning about the technologies and language features that make this possible.
I don't have one particular domain, but I've encountered domains like math (where multi-core is essential), data sort/search (where divide & conquer on multi-core is helpful) and multi-computer requirements (e.g., a requirement that a back-up station's processing power is used for something).
This depends on what language I'm working. Obviously in C#, my hands are tied with a not-yet-ready implementation of Parallel Extensions that does seem to boost performance, until you start comparing same algorithms with OpenMP (perhaps not a fair comparison). So on .NET it's going to be an easy ride with some for → Parallel.For refactorings and the like.Where things get really interesting is with C++, because the performance you can squeeze out of things like OpenMP is staggering compared to .NET. In fact, OpenMP surprised me a lot, because I didn't expect it to work so efficiently. Well, I guess its developers have had a lot of time to polish it. I also like that it is available in Visual Studio out-of-the-box, unlike TBB for which you have to pay.As for MPI, I use PureMPI.net for little home projects (I have a LAN) to fool around with computations that one machine can't quite take. I've never used MPI commercially, but I do know that MKL has some MPI-optimized functions, which might be interesting to look at for anyone needing them.
I plan to do 'frivolous computing', i.e. use extra cores for precomputation of results that might or might not be needed - RAM permitting, of course. I also intend to delve into costly algorithms and approaches that most end users' machines right now cannot handle.
As for domains not benefitting from parallellization... well, one can always find something. One thing I am concerned about is decent support in .NET, though regrettably I have given up hope that speeds similar to C++ can be attained.
I work in medical imaging and image processing.
We're handling multiple cores in much the same way we handled single cores-- we have multiple threads already in the applications we write in order to have a responsive UI.
However, because we can now, we're taking strong looks at implementing most of our image processing operations in either CUDA or OpenMP. The Intel Compiler provides a lot of good sample code for OpenMP, and is just a much more mature product than CUDA, and provides a much larger installed base, so we're probably going to go with that.
What we tend to do for expensive (ie, more than a second) operations is to fork that operation off into another process, if we can. That way, the main UI remains responsive. If we can't, or it's just far too inconvenient or slow to move that much memory around, the operation is still in a thread, and then that operation can itself spawn multiple threads.
The key for us is to make sure that we don't hit concurrency bottlenecks. We develop in .NET, which means that UI updates have to be done from an Invoke call to the UI in order to have the main thread update the UI.
Maybe I'm lazy, but really, I don't want to have to spend too much time figuring a lot of this stuff out when it comes to parallelizing things like matrix inversions and the like. A lot of really smart people have spent a lot of time making that stuff fast like nitrous, and I just want to take what they've done and call it. Something like CUDA has an interesting interface for image processing (of course, that's what it's defined for), but it's still too immature for that kind of plug-and-play programming. If I or another developer get a lot of spare time, we might give it a try. So instead, we'll just go with OpenMP to make our processing faster (and that's definitely on the development roadmap for the next few months).
So far, nothing more than more efficient compilation with make:
gmake -j
the -j option allows tasks that don't depend on one another to run in parallel.
I'm developing ASP.NET web applications. There is little possibility to use multicore directly in my code, however IIS already scales well for multiple cores/CPU's by spawning multiple worker threads/processes when under load.
We're having a lot of success with task parallelism in .NET 4 using F#. Our customers are crying out for multicore support because they don't want their n-1 cores idle!
I'm in image processing. We're taking advantage of multicore where possible by processing images in slices doled out to different threads.
I said some of this in answer to a different question (hope this is OK!): there is a concept/methodology called Flow-Based Programming (FBP) that has been around for over 30 years, and is being used to handle most of the batch processing at a major Canadian bank. It has thread-based implementations in Java and C#, although earlier implementations were fiber-based (C++ and mainframe Assembler). Most approaches to the problem of taking advantage of multicore involve trying to take a conventional single-threaded program and figure out which parts can run in parallel. FBP takes a different approach: the application is designed from the start in terms of multiple "black-box" components running asynchronously (think of a manufacturing assembly line). Since the interface between components is data streams, FBP is essentially language-independent, and therefore supports mixed-language applications, and domain-specific languages. Applications written this way have been found to be much more maintainable than conventional, single-threaded applications, and often take less elapsed time, even on single-core machines.
My graduate work is in developing concepts for doing bare-metal multicore work & teaching same in embedded systems.
I'm also working a bit with F# to bring up my high-level multiprocess-able language facilities to speed.
We create the VivaMP code analyzer for error detecting in parallel OpenMP programs.
VivaMP is a lint-like static C/C++ code analyzer meant to indicate errors in parallel programs based on OpenMP technology. VivaMP static analyzer adds much to the abilities of the existing compilers, diagnoses any parallel code which has some errors or is an eventual source of such errors. The analyzer is integrated into VisualStudio2005/2008 development environment.
VivaMP – a tool for OpenMP
32 OpenMP Traps For C++ Developers
I believe that "Cycles are an engineers' best friend".
My company provides a commercial tool for analyzing
and transforming very
large software systems in many computer languages.
"Large" means 10-30 million lines of code.
The tool is the DMS Software Reengineering Toolkit
(DMS for short).
Analyses (and even transformations) on such huge systems
take a long time: our points-to analyzer for C
code takes 90 CPU hours on an x86-64 with 16 Gb RAM.
Engineers want answers faster than that.
Consequently, we implemented DMS in PARLANSE,
a parallel programming language of our own design,
intended to harness small-scale multicore shared
memory systems.
The key ideas behind parlanse are:
a) let the programmer expose parallelism,
b) let the compiler choose which part it can realize,
c) keep the context switching to an absolute minimum.
Static partial orders over computations are
an easy to help achieve all 3; easy to say,
relatively easy to measure costs,
easy for compiler to schedule computations.
(Writing parallel quicksort with this is trivial).
Unfortunately, we did this in 1996 :-(
The last few years have finally been a vindication;
I can now get 8 core machines at Fry's for under $1K
and 24 core machines for about the same price as a small
car (and likely to drop rapidly).
The good news is that DMS is now a fairly mature,
and there are a number of key internal mechanisms
in DMS which take advantage of this, notably
an entire class of analyzers call "attribute grammars",
which we write using a domain-specific language
which is NOT parlanse. DMS compiles these
atrribute grammars into PARLANSE and then they
are executed in parallel. Our C++ front
end uses attribute grammars, and is about 100K
sloc; it is compiled into 800K SLOC of parallel
parlanse code that actually works reliably.
Now (June 2009), we are pretty busy making DMS useful, and
don't always have enough time to harness the parallelism
well. Thus the 90 hour points-to analysis.
We are working on parallelizing that, and
have reasonable hope of 10-20x speedup.
We believe that in the long run, harnessing
SMP well will make workstations far more
friendly to engineers asking hard questions.
As well they should.
Our domain logic is based heavily on a workflow engine and each workflow instance runs off the ThreadPool.
That's good enough for us.
I can now separate my main operating system from my development / install whatever I like os using vitualisation setups with Virtual PC or VMWare.
Dual core means that one CPU runs my host OS, the other runs my development OS with a decent level of performance.
Learning a functional programming language might use multiple cores... costly.
I think it's not really hard to use extra cores. There are some trivialities as web apps that does not need to have any extra care as the web server does its work running the queries in parallel. The questions are for long running algorythms (long is what you call long). These need to be split over smaller domains that does not depend each other, or synchronize the dependencies. A lot of algs can do this, but sometimes horribly different implementations needed (costs again).
So, no silver bullet until you are using imperative programming languages, sorry. Either you need skilled programmers (costly) or you need to turn to an other programming language (costly). Or you may have luck simply (web).
I'm using and programming on a Mac. Grand Central Dispatch for the win. The Ars Technica review of Snow Leopard has a lot of interesting things to say about multicore programming and where people (or at least Apple) are going with it.
I've decided to take advantage of multiple cores in an implementation of the DEFLATE algorithm. MArc Adler did something similar in C code with PIGZ (parallel gzip). I've delivered the philosophical equivalent, but in a managed code library, in DotNetZip v1.9. This is not a port of PIGZ, but a similar idea, implemented independently.
The idea behind DEFLATE is to scan a block of data, look for repeated sequences, build a "dictionary" that maps a short "code" to each of those repeated sequences, then emit a byte stream where each instance of one of the repeated sequences is replaced by a "code" from the dictionary.
Because building the dictionary is CPU intensive, DEFLATE is a perfect candidate for parallelization. i've taken a Map+Reduce type approach, where I divide the incoming uncompressed bytestreeam into a set of smaller blocks (map), say 64k each, and then compress those independently. Then I concatenate the resulting blocks together (reduce). Each 64k block is compressed independently, on its own thread, without regard for the other blocks.
On a dual-core machine, this approach compresses in about 54% of the time of the traditional serial approach. On server-class machines, with more cores available, it can potentially deliver even better results; with no server machine, I haven't tested it personally, but people tell me it's fast.
There's runtime (cpu) overhead associated to the management of multiple threads, runtime memory overhead associated to the buffers for each thead, and data overhead associated to concatenating the blocks. So this approach pays off only for larger bytestreams. In my tests, above 512k, it can pay off. Below that, it is better to use a serial approach.
DotNetZip is delivered as a library. My goal was to make all of this transparent. So the library automatically uses the extra threads when the buffer is above 512kb. There's nothing the application has to do, in order to use threads. It just works, and when threads are used, it's magically faster. I think this is a reasonable approach to take for most libbraries being consumed by applications.
It would be nice for the computer to be smart about automatically and dynamically exploiting resources on parallizable algorithms, but the reality today is that apps designers have to explicitly code the parallelization in.
I work in C# with .Net Threads.
You can combine object-oriented encapsulation with Thread management.
I've read some posts from Peter talking about a new book from Packt Publishing and I've found the following article in Packt Publishing web page:
http://www.packtpub.com/article/simplifying-parallelism-complexity-c-sharp
I've read Concurrent Programming with Windows, Joe Duffy's book. Now, I am waiting for "C# 2008 and 2005 Threaded Programming", Hillar's book - http://www.amazon.com/2008-2005-Threaded-Programming-Beginners/dp/1847197108/ref=pd_rhf_p_t_2
I agree with Szundi "No silver bullet"!
You say "For web applications it's very, very easy: ignore it. Unless you've got some code that really begs to be done in parallel you can simply write old-style single-threaded code and be happy."
I am working with Web applications and I do need to take full advantage of parallelism.
I understand your point. However, we must prepare for the multicore revolution. Ignoring it is the same than ignoring the GUI revolution in the 90's.
We are not still developing for DOS? We must tackle multicore or we'll be dead in many years.
I think this trend will first persuade some developers, and then most of them will see that parallelization is a really complex task.
I expect some design pattern to come to take care of this complexity. Not low level ones but architectural patterns which will make hard to do something wrong.
For example I expect messaging patterns to gain popularity, because it's inherently asynchronous, but you don't think about deadlock or mutex or whatever.
How does this affect your software roadmap?
It doesn't. Our (as with almost all other) business related apps run perfectly well on a single core. So long as adding more cores doesn't significantly reduce the performance of single threaded apps, we're happy
...real stories...
Like everyone else, parallel builds are the main benefit we get. The Visual Studio 2008 C# compiler doesn't seem to use more than one core though, which really sucks
What are you doing with your existing code to take advantage of multicore machines
We may look into using the .NET parallel extensions if we ever have a long-running algorithm that can be parallelized, but the odds of this actually occurring are slim. The most likely answer is that some of the developers will play around with it for interest's sake, but not much else
how will you deal with hundreds or thousands of cores?
Head -> Sand.
If your domain doesn't easily benefit from parallel computation, then explaining why is interesting, too.
The client app mostly pushes data around, the server app mostly relies on SQL server to do the heavy lifting
I'm taking advantage of multicore using C, PThreads, and a home brew implementation of Communicating Sequential Processes on an OpenVPX platform with Linux using the PREEMPT_RT patch set's scheduler. It all adds up to nearly 100% CPU utilisation across multiple OS instances with no CPU time used for data exchange between processor cards in the OpenVPX chassis, and very low latency too. Also using sFPDP to join multiple OpenVPX chassis together into a single machine. Am not using Xeon's internal DMA so as to relieve memory pressure inside CPUs (DMA still uses memory bandwidth at the expense of the CPU cores). Instead we're leaving data in place and passing ownership of it around in a CSP way (so not unlike the philosophy of .NET's task parallel data flow library).
1) Software Roadmap - we have pressure to maximise the use real estate and available power. Making the very most of the latest hardware is essential
2) Software domain - effectively Scientific Computing
3) What we're doing with existing code? Constantly breaking it apart and redistributing parts of it across threads so that each core is maxed out doing the most it possibly can without breaking out real time requirement. New hardware means quite a lot of re-thinking (faster cores can do more in the given time, don't want them to be under utilised). Not as bad as it sounds - the core routines are very modular so easily assembled into thread-sized lumps. Although we planned on taking control of thread affinity away from Linux, we've not yet managed to extract significant extra performance by doing so. Linux is pretty good at getting data and code in more or less the same place.
4) In effect already there - total machine already adds up to thousands of cores
5) Parallel computing is essential - it's a MISD system.
If that sounds like a lot of work, it is. some jobs require going whole hog on making the absolute most of available hardware and eschewing almost everything that is high level. We're finding that the total machine performance is a function of CPU memory bandwidth, not CPU core speed, L1/L2/L3 cache size.