Is GPU and SIMD likely to be implemented in .NET / Java VMs? - scala

For some time now, mainstream compute hardware has sported SIMD instructions (MMX, SSE, 3D-Now, etc) and more recently we're seeing AMD bringing 480-stream GPUs into the same die as the CPU.
Functional languages like F#, Scala and Clojure are also gaining traction, with one common attraction being how much easier concurrent programming is in these languages.
Are there any plans for the Java VM or .NET CLR to start providing access to parallel compute hardware resources, so that functional languages can mature to leverage the hardware?
It seems as though the VMs are currently the bottleneck against high performance computing, with SIMD and GPU access being delegated the 3rd party libraries and post-compilers (tidepowered.net, OpenTK, ScalaCL, Brahma, etc, etc.)
Does anyone know of any plans / roadmaps on the part of Microsoft / Oracle / Open-Source Community to being their VMs up-to-date with the new hardware and programming paradigms?
Is there a good reason why vendors are being so sluggish on the uptake?
Edit:
To Address feedback so far, it's true that GPU programming is complex and, done wrong, worsens performance. But it's well known that parallelism is the future of computing - so the crux of this question is that it doesn't help for hardware and programming languages to embrace a parallel paradigm if the runtimes sitting between the applications and the hardware don't support it... why aren't we seeing this on the VM vendor's radars / roadmaps?

you means JavaCL and ScalaCL? they both try to migrate CUDA/GPU programming to javavm

The mono runtime includes support for some SIMD instructions already - see http://docs.go-mono.com/index.aspx?link=N%3aMono.Simd
For Microsoft's implementation of the CLR you can use XNA which allows you to run shaders etc. or the accelerator library https://research.microsoft.com/en-us/projects/accelerator/ which provides an interface to running GPGPU calculations

Java has been making strong headway in the parallelism arena for some time, first with the java.util.concurrent package and now with the fork/join framework. Hopefully, in the future, languages like Clojure and Scala will provide great high level abstractions to leverage fork-join.
GPGPU programming offers significant performance gains only for very specialized problems. .Net and Java are general purpose programming languages. Plus, who wants to do CUDA-style programming in a language like Java?

Zach Tellman's Penumbra framework enables GPU programming in Clojure (both for graphics and general purpose programming).
It's somewhat experimental but I think the theoretical motivation is very sound:
Harness the GPU / specialised SIMD instructions for the serious number crunching on large data sets
Use a very high level langauge that is strong at metaprogramming / DSL definition (e.g. Clojure) to orchestrate the operations at the overall level and generate the appropriate lower-level code where needed (e.g. with generous use of macro expansions)

Related

Scalability in Scala

I am going through Scala book by Martin Odersky.
It states that Scala language is highly scalable,reason being that it allows users to add new features which can be utilised as native language support.
It has got me confused with the term 'Scalability'.
I understand that scalability means ability of a software to handle huge amount of data.
So what's the difference here?
In the context of Scala, Odersky usually means that it is scalable in the sense that it can be used for a wide range of tasks, from simple scripting to large libraries to behemoth enterprise applications.
It's good for scripting because of its type inference, relatively low verbosity (compared to Java), and functional style (which generally lends itself to more concise code).
It's good for medium size applications and libraries because of its powerful type system, which means it is possible to write code that mostly or only produces errors at compile time rather than runtime (to the extent that is possible). The Play! framework in particular is founded on this philosophy. Furthermore, Scala runs on the JVM and therefore can harness any of the many, many Java libraries out there.
And it's good for enterprise software because it compiles to JVM bytecode, which already has a great track record in enterprise software; further, the fact that it's statically typed makes the maintenance of very large codebases much easier.
Scala is also applicable to a number of other areas, making it even more "scalable": concurrency/parallelism and domain-specific languages come to mind.
Here is a presentation by Odersky, if you start at slide 6 and go forward, you'll see him explain some other uses of Scala as well.

Using SIMD instructions in application oriented to multiple platforms and OS

So, no matter how much I read about SIMD instructions, there is something basic I still can't understand properly and would, therefore, love to have some (conceptual) explanation or suggestions about.
I understand that the many SIMD implementations vary from one CPU architecture to another (MMX, SSE, SSE2, etc). However, considering that since the middle of the 2000s there seems to have been greater convergence between SIMD instructions-sets across Intel and AMD (and Apple has started used Intel), I don't get the following.
Simply put, if an application has a specific SIMD code (e.g. for a vectorized math library), would it equally run in both Intel's and AMD's (therefore in Windows and Linux computers) and also in iOS without any modification?
Or would it be required that specific code is implemented for each CPU architecture/operational system that is target by the application, in a way that different compilations of the application are given for each user type?
For Intel/AMD there can be some convergence, depending on how hard you want to push the performance envelope. iOS devices are ARM-based though, and use Neon SIMD rather than Intel/AMD's SSE/AVX, so there is no binary compatibility and only minimal compatibility at the source level (e.g. via macros or template libraries). See this question for some cross-platform solutions.

Agent Based Modeling in Modelica

Is it possible to simulate multi-agent systems in Modelica? I'm talking about a system such MASON written in Java. How easy or difficult it would be?
As I understand, Modelica is not a typical programming language, so would it be particularly helpful or will the basic design of modelica language throw any hindrance? And more importantly, how we're going to model "messaging" systems that's common in Agent-based modeling?
Modelica can simulate discrete event systems. Some libraries exist: ModelicaDEVS, ARENALib etc.
Maybe the syntax is not perfect yet for this "Messaging", but maybe the language will be improved further in this direction.
An advantage might be that real-time capable code can be created, so the agents could run in embedded systems even with hard real-time - only some of the other tools support this like Ptolemy II.
P.S. (added see first comment):
From the start Modelica was designed to create code which is capable to run in real-time. So you could take the unchanged modelica model of your agent connect IO to sensors and actuators and download it on real-time hardware (e.g. PowerPC). Your swarm of agents will then exactly fullfill the time behaviour you modeled and exist in real. Also you could have only one real agent in hardware (maybe this hardware is expensive) and simulate the interaction to all the other agents in real-time on a real-time simulator hardware using your unchanged models for that too.
This is one of the major reasons why Modelica's semantic is not that dynamic as e.g. Java. If you want to run your MASON agent on real hardware you are in trouble: you have to move to e.g. Safety Critical Java, which means that a lot of constructs of your code, but also of standard Java libraries must be rewritten or are not allowed at all. Without this you will have to live with the possibility that your agent will miss his mission and burn down the house ...

Pros and cons of using ARM assembly for Iphone development?

I'm experimenting with ARM assembly language, during this I often use inline assembly blocks in Objective-C code for my first iOS 6 small pet project. During this I wondered:
Is it reasonable at all to use ARM assembly in large commercial iOS projects?
Can I optimise some bottlenecks with it and exceed compiler in this?
What are some common guidelines or best practicies of using assembly language for iOS: vectorization, NEON, SIMD, media optimisation (image shrinking and etc.)?
The iPhone is just a small computer so the trade-offs of using assembler are exactly the same as with any other machine.
The general answer has to be no, you don't want to be using assembler.
Having said that, there are always exceptions where hand-tuned, very low level code might be faster. Make sure you use Instruments to correctly identify your bottlenecks and that you've tuned your code using higher-level techniques before you start.
Use special hardware features where it makes sense, but bear in mind that you don't need to use assembler to get the benefits of most of them.

The state of programming and compiling for multicore systems

I'm doing some research on multicore processors; specifically I'm looking at writing code for multicore processors and also compiling code for multicore processors.
I'm curious about the major problems in this field that would currently prevent a widespread adoption of programming techniques and practices to fully leverage the power of multicore architectures.
I am aware of the following efforts (some of these don't seem directly related to multicore architectures, but seem to have more to do with parallel-programming models, multi-threading, and concurrency):
Erlang (I know that Erlang includes constructs to facilitate concurrency, but I am not sure how exactly it is being leveraged for multicore architectures)
OpenMP (seems mostly related to multiprocessing and leveraging the power of clusters)
Unified Parallel C
Cilk
Intel Threading Blocks (this seems to be directly related to multicore systems; makes sense as it comes from Intel. In addition to defining certain programming-constructs, it also seems have features that tell the compiler to optimize the code for multicore architectures)
In general, from what little experience I have with multithreaded programming, I know that programming with concurrency and parallelism in mind is definitely a difficult concept. I am also aware that multithreaded programming and multicore programming are two different things. in multithreaded programming you are ensuring that the CPU does not remain idle (on a single-CPU system. As James pointed out the OS can schedule different threads to run on different cores -- but I'm more interested in describing the parallel operations from the language itself, or via the compiler). As far as I know you cannot truly do parallel operations. In multicore systems, you should be able to perform truly-parallel operations.
So it seems to me that currently the problems facing multicore programming are:
Multicore programming is a difficult concept that requires significant skill
There are no native constructs in today's programming languages that provide a good abstraction to program for a multicore environment
Other than Intel's TBB library I haven't found efforts in other programming-languages to leverage the power of multicore architectures for compilation (for example, I don't know if the Java or C# compiler optimizes the bytecode for multicore systems or even if the JIT compiler does that)
I'm interested in knowing what other problems there might be, and if there are any solutions in the works to address these problems. Links to research papers (and things of that nature) would be helpful. Thanks!
EDIT
If I had to condense my question down to one sentence, it would be this: What are the problems that face multicore programming today and what research is going on in the field to solve these problems?
UPDATE
It also seems to me that there are three levels where multicore needs to be concerned:
Language level: Constructs/concepts/frameworks that abstract parallelization and concurrency and make it easy for programmers to express the same
Compiler level: If the compiler is aware of what architecture it is compiling for, it can optimize the compiled code for that architecture.
OS level: The OS optimizes the running process and perhaps schedules different threads/processes to run on different cores.
I've searched on ACM and IEEE and have found a few papers. Most of them talk about how difficult it is to think concurrently and also how current languages don't have a proper way to express concurrency. Some have gone so far as to claim that the current model of concurrency that we have (threads) is not a good way to handle concurrency (even on multiple cores). I'm interested in hearing other views.
I'm curious about the major problems in this field that would currently prevent a widespread adoption of programming techniques and practices to fully leverage the power of multicore architectures.
Inertia. (BTW: that's pretty much the answer to all "what does prevent the widespread adoption" questions, whether that be models of parallel programming, garbage collection, type safety or fuel-efficient automobiles.)
We have known since the 1960s that the threads+locks model is fundamentally broken. By ~1980, we had about a dozen better models. And yet, the vast majority of languages that are in use today (including languages that were newly created from scratch long after 1980), offer only threads+locks.
The major problems with multicore programming is the same as writing any other concurrent applications, but whereas before it was uncommon to have multiple cpus in a computer, now it is hard to find any modern computer with only one core in it, so, to take advantage of multicore, multiple cpu architectures there are new challenges.
But, this problem is an old problem, whenever computer architectures go beyond compilers then it seems the fallback solution is to move back toward functional programming, as that programming paradigm, if strictly followed, can make very parallelizable programs, as you don't have any global mutable variables, for example.
But, not all problems can be done easily using FP, so the goal then is how to easily get other programming paradigms to be easy to use on multicores.
The first thing is that many programmers have avoided writing good mulithreaded applications, so there isn't a strongly prepared number of developers, as they learned habits that will make their coding harder to do.
But, as with most changes to the cpu, you can look at how to change the compiler, and for that you can look at Scala, Haskell, Erlang and F#.
For libraries you can look at the parallel framework extension, by MS as a way to make it easier to do concurrent programming.
It is at work, but I recently either IEEE Spectrum or IEEE Computer had articles on multicore programming issues, so look at what IEEE and ACM articles have been written on these issues, to get more ideas as to what is being looked at.
I think the biggest impediment will be the difficulty to get programmers to change their language as FP is very different than OOP.
One place for research besides developing languages that will work well this way, is how to handle multiple threads accessing memory, but, as with much in this area, Haskell seems to be at the forefront in testing ideas for this, so you can look at what is going on with Haskell.
Ultimately there will be new languages, and it may be that we have DSLs to help abstract the developer more, but how to educate programmers on this will be a challenge.
UPDATE:
You may find Chapter 24. Concurrent and multicore programming of interest, http://book.realworldhaskell.org/read/concurrent-and-multicore-programming.html
One of the answers mentioned the Parallel Extensions for the .NET Framework and since you mentioned C#, it's definitely something I would investigate. Microsoft has done something interesting things there, though I have to think many of their efforts seem more suited for language enhancements in C# than a separate and distinct library for concurrent programming. But I think their efforts are worth applauding and respect that we're early here. (Disclaimer: I used to be the marketing director for Visual Studio about 3 years ago)
The Intel Thread Building Blocks are also quite interesting (Intel recently released a new version, and I'm excited to head down to Intel Developer Forum next week to learn more about how to use it properly).
Lastly, I work for Corensic, a software quality startup in Seattle. We've got a tool called Jinx that is designed to detect concurrency errors in your code. A 30-day trial edition is available for Windows and Linux, so you might want to check it out. (www.corensic.com)
In a nutshell, Jinx is a very thin hypervisor that, when activated, slips in between the processor and operating system. Jinx then intelligently takes slices of execution and runs simulations of various thread timings to look for bugs. When we find a particular thread timing that will cause a bug to happen, we make that timing "reality" on your machine (e.g., if you're using Visual Studio, the debugger will stop at that point). We then point out the area in your code where the bug was caused. There are no false positives with Jinx. When it detects a bug, it's definitely a bug.
Jinx works on Linux and Windows, and in both native and managed code. It is language and application platform agnostic and can work with all your existing tools.
If you check it out, please send us feedback on what works and doesn't work. We've been running Jinx on some big open source projects and already are seeing situations where Jinx can find bugs 50-100 times faster than simply stress testing code.
The bottleneck of any high-performance application (written in C or C++) designed to make efficient use of more than one processor/core is the memory system (caches and RAM). A single core usually saturates the memory system with its reads and writes so it is easy to see why adding extra cores and threads causes an application to run slower. If a queue of people can pass through a door one a time, adding extra queues will not only clog the door but also make the passage of any one individual through the door less efficient.
The key to any multi-core application is optimization of and economizing on memory accesses. This means structuring data and code to work as much as possible inside their own caches where they don't disturb the other cores with acceses to the common cache (L3) or RAM. Once in a while a core needs to venture there but the trick is to reduce those situations as much as possible. In particular, data needs to be structured around and adapted to cache lines and their sizes (currently 64 bytes) and code needs to be compact and not call and jump all over the place which also disrupts pipelines.
My experience is that efficient solutions are unique to the application in question. The generic guidelines (above) are a basis on which to construct code but the tweak changes resulting from profiling conclusions will not be obvious to those who were not themselves involved in the optimizing work.
Look up fork/join frameworks and work-stealing runtimes. Two names for the same, or at least related, approaches, which is to recursively subdivide large tasks into lightweight units, such that all available parallelism is exploited, without having to know in advance how much parallelism there is. The idea is that it should run at serial speed on a uniprocessor, but get a linear speedup with multiple cores.
Sort of a horizontal analogue of cache-oblivious algorithms if you look at it right.
But i'd say the main problem facing multicore programming is that the great majority of computations remain stubbornly serial. There's just no way to throw multiple cores at those computations and make them stick.