How to mix Coarray and MPI code in Fortran - fortran

I would like to combine Fortran coarray with MPI inside my code. I plan to use third party softwave(HYPRE), which used MPI, for linear system solvers. For the rest of my work, I want to use the Fortran coarray (OpenCoarrays). I've already searched for a solution on Internet. But there isn't any clues about how to make it work. I wonder is that possible to mix Fortran coarray and MPI. If yes, should I use the OpenCoarrays or MPI wrapper compilers

OpenCoarrays sits on top of MPI-3 RMA (at least by default; I don’t recall the latest status of the GASNet port) so this should work, even if neither standard guarantees this. You’ll be using process-parallel execution and they should interoperate fine.
Intel Fortran also uses MPI for coarrays. Cray Fortran coarrays use DMAPP, which is compatible with MPI. Thus, the interoperability you want should cover all the widely available implementations.
In all cases, there may be some implementation quirks, particularly with respect to initialization and termination. You may find that you can’t finalize MPI until all your coarrays are deallocated, for example.
I’m sure the developers of OpenCoarrays would appreciate big reports on this topic if you have problems.

Related

Explanation of CUDA C and C++

Can anyone give me a good explanation as to the nature of CUDA C and C++? As I understand it, CUDA is supposed to be C with NVIDIA's GPU libraries. As of right now CUDA C supports some C++ features but not others.
What is NVIDIA's plan? Are they going to build upon C and add their own libraries (e.g. Thrust vs. STL) that parallel those of C++? Are they eventually going to support all of C++? Is it bad to use C++ headers in a .cu file?
CUDA C is a programming language with C syntax. Conceptually it is quite different from C.
The problem it is trying to solve is coding multiple (similar) instruction streams for multiple processors.
CUDA offers more than Single Instruction Multiple Data (SIMD) vector processing, but data streams >> instruction streams, or there is much less benefit.
CUDA gives some mechanisms to do that, and hides some of the complexity.
CUDA is not optimised for multiple diverse instruction streams like a multi-core x86.
CUDA is not limited to a single instruction stream like x86 vector instructions, or limited to specific data types like x86 vector instructions.
CUDA supports 'loops' which can be executed in parallel. This is its most critical feature. The CUDA system will partition the execution of 'loops', and run the 'loop' body simultaneously across an array of identical processors, while providing some of the illusion of a normal sequential loop (specifically CUDA manages the loop "index"). The developer needs to be aware of the GPU machine structure to write 'loops' effectively, but almost all of the management is handled by the CUDA run-time. The effect is hundreds (or even thousands) of 'loops' complete in the same time as one 'loop'.
CUDA supports what looks like if branches. Only processors running code which match the if test can be active, so a subset of processors will be active for each 'branch' of the if test. As an example this if... else if ... else ..., has three branches. Each processor will execute only one branch, and be 're-synched' ready to move on with the rest of the processors when the if is complete. It may be that some of the branch conditions are not matched by any processor. So there is no need to execute that branch (for that example, three branches is the worst case). Then only one or two branches are executed sequentially, completing the whole if more quickly.
There is no 'magic'. The programmer must be aware that the code will be run on a CUDA device, and write code consciously for it.
CUDA does not take old C/C++ code and auto-magically run the computation across an array of processors. CUDA can compile and run ordinary C and much of C++ sequentially, but there is very little (nothing?) to be gained by that because it will run sequentially, and more slowly than a modern CPU. This means the code in some libraries is not (yet) a good match with CUDA capabilities. A CUDA program could operate on multi-kByte bit-vectors simultaneously. CUDA isn't able to auto-magically convert existing sequential C/C++ library code into something which would do that.
CUDA does provides a relatively straightforward way to write code, using familiar C/C++ syntax, adds a few extra concepts, and generates code which will run across an array of processors. It has the potential to give much more than 10x speedup vs e.g. multi-core x86.
Edit - Plans: I do not work for NVIDIA
For the very best performance CUDA wants information at compile time.
So template mechanisms are the most useful because it gives the developer a way to say things at compile time, which the CUDA compiler could use. As a simple example, if a matrix is defined (instantiated) at compile time to be 2D and 4 x 8, then the CUDA compiler can work with that to organise the program across the processors. If that size is dynamic, and changes while the program is running, it is much harder for the compiler or run-time system to do a very efficient job.
EDIT:
CUDA has class and function templates.
I apologise if people read this as saying CUDA does not. I agree I was not clear.
I believe the CUDA GPU-side implementation of templates is not complete w.r.t. C++.
User harrism has commented that my answer is misleading. harrism works for NVIDIA, so I will wait for advice. Hopefully this is already clearer.
The hardest stuff to do efficiently across multiple processors is dynamic branching down many alternate paths because that effectively serialises the code; in the worst case only one processor can execute at a time, which wastes the benefit of a GPU. So virtual functions seem to be very hard to do well.
There are some very smart whole-program-analysis tools which can deduce much more type information than the developer might understand. Existing tools might deduce enough to eliminate virtual functions, and hence move analysis of branching to compile time. There are also techniques for instrumenting program execution which feeds directly back into recompilation of programs which might reach better branching decisions.
AFAIK (modulo feedback) the CUDA compiler is not yet state-of-the-art in these areas.
(IMHO it is worth a few days for anyone interested, with a CUDA or OpenCL-capable system, to investigate them, and do some experiments. I also think, for people interested in these areas, it is well worth the effort to experiment with Haskell, and have a look at Data Parallel Haskell)
CUDA is a platform (architecture, programming model, assembly virtual machine, compilation tools, etc.), not just a single programming language. CUDA C is just one of a number of language systems built on this platform (CUDA C, C++, CUDA Fortran, PyCUDA, are others.)
CUDA C++
Currently CUDA C++ supports the subset of C++ described in Appendix D ("C/C++ Language Support") of the CUDA C Programming Guide.
To name a few:
Classes
__device__ member functions (including constructors and destructors)
Inheritance / derived classes
virtual functions
class and function templates
operators and overloading
functor classes
Edit: As of CUDA 7.0, CUDA C++ includes support for most language features of the C++11 standard in __device__ code (code that runs on the GPU), including auto, lambda expressions, range-based for loops, initializer lists, static assert, and more.
Examples and specific limitations are also detailed in the same appendix linked above. As a very mature example of C++ usage with CUDA, I recommend checking out Thrust.
Future Plans
(Disclosure: I work for NVIDIA.)
I can't be explicit about future releases and timing, but I can illustrate the trend that almost every release of CUDA has added additional language features to get CUDA C++ support to its current (In my opinion very useful) state. We plan to continue this trend in improving support for C++, but naturally we prioritize features that are useful and performant on a massively parallel computational architecture (GPU).
Not realized by many, CUDA is actually two new programming languages, both derived from C++. One is for writing code that runs on GPUs and is a subset of C++. Its function is similar to HLSL (DirectX) or Cg (OpenGL) but with more features and compatibility with C++. Various GPGPU/SIMT/performance-related concerns apply to it that I need not mention. The other is the so-called "Runtime API," which is hardly an "API" in the traditional sense. The Runtime API is used to write code that runs on the host CPU. It is a superset of C++ and makes it much easier to link to and launch GPU code. It requires the NVCC pre-compiler which then calls the platform's C++ compiler. By contrast, the Driver API (and OpenCL) is a pure, standard C library, and is much more verbose to use (while offering few additional features).
Creating a new host-side programming language was a bold move on NVIDIA's part. It makes getting started with CUDA easier and writing code more elegant. However, truly brilliant was not marketing it as a new language.
Sometimes you hear that CUDA would be C and C++, but I don't think it is, for the simple reason that this impossible. To cite from their programming guide:
For the host code, nvcc supports whatever part of the C++ ISO/IEC
14882:2003 specification the host c++ compiler supports.
For the device code, nvcc supports the features illustrated in Section
D.1 with some restrictions described in Section D.2; it does not
support run time type information (RTTI), exception handling, and the
C++ Standard Library.
As I can see, it only refers to C++, and only supports C where this happens to be in the intersection of C and C++. So better think of it as C++ with extensions for the device part rather than C. That avoids you a lot of headaches if you are used to C.
What is NVIDIA's plan?
I believe the general trend is that CUDA and OpenCL are regarded as too low level techniques for many applications. Right now, Nvidia is investing heavily into OpenACC which could roughly be described as OpenMP for GPUs. It follows a declarative approach and tackles the problem of GPU parallelization at a much higher level. So that is my totally subjective impression of what Nvidia's plan is.

Parallel computing using computers in LAN?

I'm writing a C++ project to solve the Travelling salesman problem using genetic algorithms. Naturally, I'd like to make it faster using a bunch (about 40) computers in the same LAN. The computers are all running Windows XP... So, the question is what are the ways to parallelize it using the given equipment
Update:
You've helped me to narrow my choices down to mpich an Open MPI, so the only question left is should I use boost MPI wrappers for them? Also, can you recommend a tutorial for mpich/OpenMPI?
You'll need to use some form of communication to synchronize the processes between the systems. A common API for this type of application is to use Message Passing Interface (MPI). There are quite a few implementations for MPI/MPI-2 that work on Windows XP.
MPI is one of the most used standards for such kind of projects. There are many implementations of this standard from different vendors. I think you should consider two of them:
mpich http://www.mcs.anl.gov/research/projects/mpich2/
Intel MPI http://www.intel.com/go/mpi/
Another interesting way is to use OpenMP. There are Cluster OpenMP implementations. This solution could be easier to implement, but not so scalable as MPI in some cases.

Data parallel libraries in C/C++

I have a C# prototype that is heavily data parallel, and I've had extremely successful order of magnitude speed ups using the Parallel.For construct in .NETv4. Now I need to write the program in native code, and I'm wondering what my options are. I would prefer something somewhat portable across various operating systems and compilers, but if push comes to shove, I can go with something Windows/VC++.
I've looked at OpenMP, which looks interesting, but I have no idea whether this is looked upon as a good solution by other programmers. I'd also love suggestions for other portable libraries I can look at.
If you're happy with the level of parallelism you're getting from Parallel.For, OpenMP is probably a pretty good solution for you -- it does roughly the same kinds of things. There's also work and research being done on parallelized implementations of the algorithms in the standard library. Code that uses the standard algorithms heavily can gain from this with even less work.
Some of this is done using OpenMP, while other is using hand-written code to (attempt to) provide greater benefits. In the long term, we'll probably see more of the latter, but for now, the OpenMP route is probably a bit more practical.
If you're using Parallel.For in .Net 4.0, you should also look at the Parallel Pattern Library in VS2010 for C++; many things are similar and it is library based.
If you need to run on another platform besides Windows the PPL is also implemented in Intel's Thread Building Blocks which has an open source version.
Regarding the other comment on similarities / differences vs. .Net 4.0, the parallel loops in the PPL (parallel_for, parallel_for_each, parallel_invoke) are virtually identical to the .Net 4.0 loops. The task model is slightly different, but should be straightforward.
You should check out Intel's Thread Building Blocks. Visual Studio 2010 also offers a Concurrency Runtime native. The .NET 4.0 libraries and the ConcRT are designed very similarly, if not identically.
if you want something that is versatile, as in portable across various OS and environments, it would be very difficult to not to consider Java. And they are very similar to C# so it be a very easy transition.
Unless you want to pull out your ninja scalpel and wanting to make your code extremely efficient, I would say java over VC++ or C++.

Developing embedded software library, C or C++?

I'm in the process of developing a software library to be used for embedded systems like an ARM chip or a TI DSP (for mostly embedded systems, but it would also be nice if it could also be used in a PC environment). Obviously this is a pretty broad range of target systems, so being able to easily port to different systems is a priority.The library will be used for interfacing with a specific hardware and running some algorithms.
I am thinking C++ is the best option, over C, because it is much easier to maintain and read. I think the additional overhead is worth it for being able to work in the object oriented paradigm. If I was writing for a very specific system, I would work in C but this is not the case.
I'm assuming that these days most compilers for popular embedded systems can handle C++. Is this correct?
Is there any other factors I should consider? Is my line of thinking correct?
If portability is very important for you, especially on an embedded system, then C is certainly a better option than C++. While C++ compilers on embedded platforms are catching up, there's simply no match for the widespread use of C, for which any self-respecting platform has a compliant compiler.
Moreover, I don't think C is inferior to C++ where it comes to interfacing hardware. The amount of abstraction is sufficiently low (i.e. no deep class hierarchies) to make C just as good an option.
There is certainly good support of C++ for ARM. ARM have their own compiler and g++ can also generate EABI compliant ARM code. When it comes to the DSPs, you will have to look at their toolchain to decide what you are going to do. Be aware that the library that comes with a DSP may well not implement the full C or C++ standard library.
C++ is suitable for low-level embedded development and is used in the SymbianOS Kernel. Having said that, you should keep things as simple as possible.
Avoid exceptions which may demand more library support than what is present (therefore use new (std::nothrow) Foo instead of new Foo).
Avoid memory allocations as much as possible and do them as early as possible.
Avoid complex patterns.
Be aware that templates can bloat your code.
I have seen many complaints that C++ is "bloated" and inappropriate for embedded systems.
However, in an interview with Stroustrup and Sutter, Bjarne Stroustrup mentioned that he'd seen heavily templated C++ code going into (IIRC) the braking systems of BMWs, as well as in missile guidance systems for fighter aircraft.
What I take away from this is that experts of the language can generate sophisticated, efficient code in C++ that is most certainly suitable for embedded systems. However, a "C With Classes"[1] programmer that does not know the language inside out will generate bloated code that is inappropriate.
The question boils down to, as always: in which language can your team deliver the best product?
[1] I know that sounds somewhat derogatory, but let me say that I know an awful lot of these guys, and they churn out an awful lot of relatively simple code that gets the job done.
C++ compilers for embedded platforms are much closer to 83's C with classes than 98's C++ standard, let alone C++0x. For instance, some platform we use still compile with a special version of gcc made from gcc-2.95!
This means that your library interface will not be able to provide interfaces with containers/iterators, streams, or such advanced C++ features. You'll have to stick with simple C++ classes, that can very easily be expressed as a C interface with a pointer to a structure as first parameter.
This also means that within your library, you won't be able to use templates to their full power. If you want portability, you will still be restricted to generic containers use of templates, which is, I'm sure you'll admit, only a very tiny part of C++ templates power.
C++ has little or no overhead compared to C if used properly in an embedded environment. C++ has many advantages for information hiding, OO, etc. If your embedded processor is supported by gcc in C then chances are it will also be supported with C++.
On the PC, C++ isn't a problem at all -- high quality compilers are extremely widespread and almost every C compiler is directly associated with a C++ compiler that's quite good, though there are a few exceptions such as lcc and the newly revived pcc.
Larger embedded systems like those based on the ARM are generally quite similar to desktop systems in terms of tool chain availability. In fact, many of the same tools available for desktop machines can also generate code to run on ARM-based machines (e.g., lots of them use ports of gcc/g++). There's less variety for TI DSPs (and a greater emphasis on quality of generated code than source code features), but there are still at least a couple of respectable C++ compilers available.
If you want to work with smaller embedded systems, the situation changes in a hurry. If you want to be able to target something like a PIC or an AVR, C++ isn't really much of an option. In theory, you could get (for example) Comeau to produce a custom port that generated code you could compile on that target's C compiler -- but chances are pretty good that even if you did, it wouldn't work out very well. These systems are really just too limitated (especially on memory size) for C++ to fit them well.
Depending on what your intended use is for the library, I think I'd suggest implementing it first as C - but the design should keep in mind how it would be incorporated into a C++ design. Then implement C++ classes on top of and/or along side of the C implementation (there's no reason this step cannot be done concurrently with the first). If your C design is done with a C++ design in mind, it's likely to be as clean, readable and maintainable as the C++ design would be. This is somewhat more work, but I think you'll end up with a library that's useful in more situations.
While you'll find C++ used more and more on various embedded projects, there are still many that restrict themselves to C (and I'd guess this is more often the case than not) - regardless of whether or not the tools support C++. It would be a shame to have a nice library of routines that you could bring to a new project you're working on, but be unable to use them because C++ isn't being used on that particular project.
In general, it's much easier to use a well-designed C library from C++ than the other way around. I've taken this approach with several sets of code including parsing Intel Hex files, a simple command parser, manipulating synchronization objects, FSM frameworks, etc. I'm planning on doing a simple XML parser at some point.
Here's an entirely different C++-vs-C argument: stable ABIs. If your library exports a C ABI, it can be compiled with any compiler that works on the system, because C ABIs are generally platform standards. If your library exports a C++ ABI, it can only be compiled with a matching compiler -- because C++ ABIs are usually not platform standards, and often differ from compiler to compiler and even version to version.
Interestingly, one of the rare exceptions to this is ARM; there's an ARM C++ ABI specification, and all compliant ARM compilers follow it. This is not true on x86; on x86, you're lucky if a C++ library compiled with a 4.1 version of GCC will link correctly with an application compiled with GCC 4.4, and don't even ask about 3.4.6.
Even if you export a C ABI, you can have problems. If your library uses C++ internally, it will then link to libstdc++ for things in the C++ std:: namespace. If your user compiles a C++ application that uses your library, they'll also link to libstdc++ -- and so the overall application gets linked to libstdc++ twice, and their libstdc++ may not be compatible with your libstdc++, which can (or so I understand) lead to odd errors from the intersection of the two. Considerably less likely, but still possible.
All of these arguments only apply because you're writing a library, and they're not showstoppers. But they are things to be aware of.

Adding MPI support to a C++ program

I have a program that is been implemented in C++ which I now want to add MPI support. There is an MPI binding for C++, with namespace MPI and everything.
In my case I have a specific object that is suitable to be the parallelized process into the cluster.
My questions are:
Has anyone done something like this before? Can I get some advices on how best to implement this?
How do I initialize MPI inside the constructor? After initializing MPI inside the constructor of the Class, will all the intermediate calls be parallelized too?
For example:
MyClass obj;
x = x; //this will be parallelized ?
onj.calc();
y = x++; //this will be parallelized ?
z = obj.result();
MPI doesn't parallelize anything automatically, it only gives you an interface for sending data between nodes. Your code is written and runs as usual sequential code independently on each node and every once in a while you send data to some other node or try to receive data from some other node.
I would really recommend picking up the Gropp MPI Book, it really helps for basic MPI!
Chiming in on an old thread, I found OpenMPI and Boost::MPI nice to work with. The object oriented design of the library may be a bit bolted on, but I found it much nicer to work with than pure MPI, especially with auto-serialization of many types and a rather extendable interface for gather/reduce functions as well as serialization of user types.
As background information:
Most applications that use MPI are written in Fortran or C. Every major implementation of MPI is written in C.
Support for MPI C++ bindings is sketchy at best: Some of the MPI Datatypes are not available (e.g. MPI_DOUBLE), there are issues with I/O and the order that headers are included in the source files. There are issues of name mangling if the MPI library was built with C and the application is built with Fortran or C++. mpich2 FAQ has entries to help work through these issues. I am less familiar with Open MPI and it's particular behavior with Fortran and C++.
For your specific questions:
I think that you have a fundamental mis-understanding about what MPI is and is not, and how application code should interact with the MPI libraries.
Parallel Programming with MPI is an excellent reference for learning to program with MPI. The code examples are in C, and most of the MPI API's are shown in an example. I highly recommend that you work through this book to learn what parallel programing is and is not.