|
forums.silverfrost.com Welcome to the Silverfrost forums
|
View previous topic :: View next topic |
Author |
Message |
DanRRight
Joined: 10 Mar 2008 Posts: 2866 Location: South Pole, Antarctica
|
Posted: Sat Nov 25, 2023 12:02 am Post subject: Supercomputing-23 |
|
|
Good read to stay in trend
https://www.hpcwire.com/2023/11/20/supercomputing-2023-odds-and-ends-from-the-show/
Will note that in supercomputing world which is 100% Linux based there is nothing such convenient as SDBG/SDBG64 and Clearwin with OpenGL. People still debug grandfather's way planting Print* into the possible offending place of code.
If Silverfrost offered this to supercomputers with MPI and multithreading debugging that would change the entire landscape of scientific simulations and this compiler would be used for decades to come |
|
Back to top |
|
|
JohnCampbell
Joined: 16 Feb 2006 Posts: 2584 Location: Sydney
|
Posted: Sun Nov 26, 2023 2:50 am Post subject: |
|
|
Dan,
There is a lot of work to achieve this.
OpenMP is a significant step, but does not include MPI.
OpenMP Ver 4 would give some good functionality, but Ver 5+ is a significant increase in complexity, which is associated with GPU off-loading.
I am not familiar with MPI usage at all.
The other significant step is increased use of AVX instructions.
FTN95 main strengths are its diagnostics, its program development aids and Clearwin+.
Can you idenify any group who could assist to merge FTN95 with OpenMP and probably Lunix ? |
|
Back to top |
|
|
DanRRight
Joined: 10 Mar 2008 Posts: 2866 Location: South Pole, Antarctica
|
Posted: Mon Nov 27, 2023 11:01 am Post subject: |
|
|
John, It will be almost no sense in using OpenMP or MPI without SDBG64 / Plato. Debugger must stop/continue specific thread, launch on error exactly where it happened etc. Without that any Fortran is not better than other dumb languages lacking graphical debuggers. And this is why people abandon Fortran to C/C++ as they do not see software support and debugging advantages, development in both is hard, painful and annoying, but at least C/C++ has applications outside of science and engineering.
People do not know what good Fortran debugger is and that they could work with the charm. Do not know about it probably even 99% of users of this compiler. My impression is that only few in this forum use SDBG, others work like blind-deaf-muted.
The only who can help here are ... Paul and Robert and Silverfrost. Adding more complexity without debugging tools has no sense. This compiler will lose its major advantages over the others
Power users like yourself using gFortran, mecej4 (saw his OpenMP examples in the Intel Forum), some others who use parallelism outside FTN95 should try to convince this company to implement industry standard parallelization. Currently FTN95 supports few multithreading methods:
- parallel algebra libraries (compiled with other compilers and linked to FTN95 via DLL)
- some launching threads functions (Silverfrost proprietary)
- FTN95 for NET multythread functions (Silverfrost proprietary)
but
- OpenMP not supported
- MPI not supported
- CUDA not supported...
If we can not use multithreaded software built by other compilers without total rebuild we lose major advantage of Fortran over other languages - large legacy code base.
I do not use OpenMP, i use MPI both on PC and supercomputers. Differences with MPI and OpenMP are that first one can also work on different nodes (supercomputers and most servers consist of many of them, nodes locally share own memory and other resources and are connected via fast switches) while OpenMP only on one node like all our PCs or HEDT ( high end desktops).
HEDT become more and more powerful too, and even currently are not bad either. AMD Threadripper for example has 64 cores, soon will be 96 and 128. Ten years ago this power was delivered by 200-300-core supercomputer. And if you add GPU then you will get a whole 1000-core equivalent. Even AMD 7950X3D usual consumer processor for PC is not bad. It crushes some HEDT, see recent tests on Anandtech, it has large cache and also supports AVX512 and larger frequencies. Hopefully AMD also cleaned some its initial defects.
Very good simple read about MPI and OpenMP is here: (very well written!)
https://curc.readthedocs.io/en/latest/programming/MPI-Fortran.html
Just do not use Code: | module load gcc
module load openmpi
or
module load intel
module load impi | for gFortran and Intel respectively on your Linux PC because compilers and MPI you will install permanently at the beginning, these commands are mostly for supercomputers (unless you install such boring version of Linux on your PC) |
|
Back to top |
|
|
|
|
You cannot post new topics in this forum You cannot reply to topics in this forum You cannot edit your posts in this forum You cannot delete your posts in this forum You cannot vote in polls in this forum
|
Powered by phpBB © 2001, 2005 phpBB Group
|