Back to Table of contents

Primeur weekly 2014-11-20

Focus

OpenMP to launch code challenge, taking parallel open source code to the next level ...

Exascale supercomputing

Bull, an Atos company, takes the lead in the extreme computing race ...

En route to processing power that rivals the human brain ...

SGI extends high performance computing innovation on path to exascale ...

EuroFlash

Numascale, Supermicro, and AMD announce the world's largest shared memory system to date ...

The University of Tromsø expands use of Asetek's liquid cooling ...

OCCIGEN: Pushing the frontiers of science and technology ...

Eurotech introduces HiVe, the ultimate accelerated computing solution ...

King Mongkut's University of Technology Thonburi acquires a NumaQ system for high performance in-memory analytics computations ...

National High Performance Computing Centre, Shenzhen University acquires Numascale shared memory system ...

Altair partners with South Africa's Centre for High Performance Computing ...

Czech Republic National Supercomputer Center selects powerful SGI supercomputer to drive advanced scientific research ...

USFlash

Altera and IBM unveil FPGA-accelerated POWER systems with coherent shared memory ...

Data Vortex to introduce their first scalable fine-grained gather-scatter parallel computer at Supercomputing 2014 ...

Kitware works with NVIDIA to drive massive acceleration in HPC visualization ...

Lenovo showcases high-performance computing innovations at Supercomputing 2014 ...

New Orleans becomes home to fastest Internet hub in the world ...

PGI high performance computing compilers coming to IBM POWER systems ...

Mellanox InfiniBand leads TOP500 as most used interconnect technology for high-performance computing for the 2nd consecutive time ...

Mellanox introduces programmable network adapter product line with application acceleration engine ...

Mellanox delivers the world's fastest EDR 100Gb/s InfiniBand switch with latency less than 90 nanosecond ...

Minnesota Supercomputing Institute selects Mellanox EDR 100Gb/s InfiniBand for large-scale supercomputer ...

Rogue Wave reveals new releases and a presentation theater at SC14 ...

ANSYS and NCSA achieve supercomputing milestone ...

Seagate announces reseller agreement with SGI ...

OpenMP to launch code challenge, taking parallel open source code to the next level

19 Nov 2014 New Orleans - At SC14 in New Orleans,Primeur Magazinehad the opportunity to interview Michael Wong, CEO of the OpenMP Architecture Review Board (ARB). Michael Wong toldPrimeur Magazinethat OpenMP up to now has been a kind of de facto standard in about being the most capable parallel language. The OpenMP team wants to make sure that it's easy to take any code from any other parallel language and convert it into OpnMP. The team wants to show that not all on that source to source conversion point of view can be done and if it can be done, then it is something that we need when there is a need to improve. Ultimately, everybody wins. OpenMP hopes that ultimately the industry wins. Because there are a lot of different parallel languages out there, it would be good if we can see that, stated Michael Wong. Certainly, there is an amount of misconception that OpenMP is still the old open OpenMP. OpenMP changed its mission statement this year, essentially broadening its mandate to cover all forms of parallelism.

So with that in mind and the fact that I actually think OpenMP is probably the only language that is capable of addressing all the MIPS on your machine, I actually am very active in other committees as well too, Michael Wong added. Maybe, languages are missing certain aspects of letting you access your vector unit, your GPU accelerator units, so I think OpenMP actually is in a pretty good position. On top of that, it actually works with three different general purpose languages as well.

Michael Wong hopes that the code challenge will be successful, taking open source code to the level that people can put on to GitHub. One of the OpenMP experts, we have, is basically exquisite, covering every field within our group: Open ACC openCL, Cuda on C++. I am on the C++ standard committee and actually wrote a lot of the wording that allows C++ to be parallel. So we have enough experience with HPC. We should be able to get an expert on board to do the conversion and show that ultimately, OpenMP is the most capable parallel language out there, Michael Wong added.

Primeur Magazine:So you expect that people send in their codes and then you look at them and make them parallel in OpnMP?

Michael Wong:Yes, so people would give us a note on the OpenMP forum and they would post-to-code on GitHub. We have to ask them if the license is open, otherwise we cannot look at it. Right now, we are starting the challenge so we are finding our way through it, how long it takes and all that. We will give a formal response on the OpenMP forum about the results. It is unlikely we are going to convert the whole code. Mostly, people will just be interested in a particular region of the code: the parallel regions they want to convert. That is what our experts will be looking at.

It is good for us and it's good for the industry. It is good for us because it will help us improve the language to see if there is something missing. It is good for the industry because there will coming out of it an improved OpenMP. And also the other languages will improve. So I think it's good for everybody.

Primeur Magazine:Is there a closing date for the challenge?

Michael Wong:I think this is a challenge that comes out at the Supercomputer Conference, SC14, this year. To be honest we would really do this anyway behind the scene. What we are doing is formalizing the process and have a little bit of fun with it. So if there was a closing date, that would be the next Supercomputer Conference, at which point we might modify the challenge, or even add on top of it. That is our thinking at this point.

Primeur Magazine:Are not you afraid that people will tend to forget it until it's too late?

Michael Wong:That is the way we are promoting it, with a lot of PR. It will be on our website. We will be hosting an OpenMP User Conference next year in Aachen. That will be a big draw from all the users to come in. If there is code that we need to look at, we will talk about that at the conference which is going to be in September 2015. So, that is one of the keystones of the conference: look at the results of this challenge and see what it looks like.

Primeur Magazine:Did you already have people that responded to the challenge?

Michael Wong:We just opened the challenge yesterday. The people that know have told that they will send us something. But I can tell you that we have been looking already at Open ACC to OpenMP code, we have been looking at CL code to OpenMP, Cuda to OpenMP. So, some of this might not be public because some of our people are in both committees for instance. So nothing public up till now we can say, but for this it is mostly a public open challenge.

Primeur Magazine:There is also a new version of the OpenMP standard?

Michael Wong:Yes, TR3 is the new version we have been working on. This is normal now in OpenMP with new product releases every year. Whether it is an official release or whether it is a technical report. So TR3 contains specifics about enhancing accelerators. This will give people an early look ahead. There can be early implementations so that by the time OPenMP 4 comes out in 2015 at SC15, it will be really solid, because we are not going to stop. We are adding to it, but this TR3 is a statement announcing this is what we pretty much are going to work out with for OpenMP 4.1.

Primeur Magazine:So, it is a kind of preview. People can look at it and comment?

Michael Wong:Absolutely. Developers can start implementing it. Users can start giving us feedback. As with any technical report, there is a 60 days comments and feedback period, and then, we will probably will close it and it will probably continue to evolve. It could be that we change it or we could standardize it or standardize only parts of it, or change components of it. That is typically how technical reports are being dealt with.

Primeur magazine:The other report published is an API report. What is the status of that report?

Michael Wong:TR2 is out 6 months now. The report is edited by John Mellor from Rice University and a researcher from IBM, Alexandre Eichenberger, and others. It essentially specifies the interface to tools. Now, the initial report specifies very tightly what interface is needed to debuggers for instance. For that, we've had people from Etnus, TotalView being involved in the design of that whole idea. The idea is that there are toolkits being implemented at Rice University and IBM, as well as other locations. They will implement an interface which allows people to actively be able to get information about the OpenMP parallel region status. That's actually quite far. Again at some point that it could also turn into a standard. Most likely by 5.0, because it is a big enough package, so that you cannot just do it in a "dot-release". You probably have to encapsulate it in part into 5.0.

By then people will already have usage experience, implementation experience, and certainly by then it will be good enough to be standardized. That is the status. It has implementation and it is moving forward.

The report has already been out since February/March and the comment period expired some time ago. But to be honest, we are always willing to accept further comments. The comment phase has been given a bit of an artificial cut-off date, but really at any time people can send comments through the form to the editors. The names of the editors and their e-mails are on the report so you can send them notes with suggestions.

Primeur Magazine:Can you tell a bit about the conference?

Michael Wong:It will be a big major conference for the users of OpenMP. We are looking to get lots of good talks. Most of them will be given by the users themselves. Some of them will be given by committee members to tell people about the rationale for a certain language design. So, it is mostly a user driven conference. I think we have been needing something like this since a long, long time. There are a lot of users of OpenMP.

I think we will have a wonderful conference looking for something like 100 people, maybe 200, hosted at a nice big place at Aachen in Germany. There will be an OpenMP conference and a tutorial, and following that, there will be the face-to-face language meeting. It is a co-located event: a multi-conference. So you can stay between 2 and 5 days.

Primeur Magazine:Thanks for the interview.
Ad Emmen

Back to Table of contents

Primeur weekly 2014-11-20

Focus

OpenMP to launch code challenge, taking parallel open source code to the next level ...

Exascale supercomputing

Bull, an Atos company, takes the lead in the extreme computing race ...

En route to processing power that rivals the human brain ...

SGI extends high performance computing innovation on path to exascale ...

EuroFlash

Numascale, Supermicro, and AMD announce the world's largest shared memory system to date ...

The University of Tromsø expands use of Asetek's liquid cooling ...

OCCIGEN: Pushing the frontiers of science and technology ...

Eurotech introduces HiVe, the ultimate accelerated computing solution ...

King Mongkut's University of Technology Thonburi acquires a NumaQ system for high performance in-memory analytics computations ...

National High Performance Computing Centre, Shenzhen University acquires Numascale shared memory system ...

Altair partners with South Africa's Centre for High Performance Computing ...

Czech Republic National Supercomputer Center selects powerful SGI supercomputer to drive advanced scientific research ...

USFlash

Altera and IBM unveil FPGA-accelerated POWER systems with coherent shared memory ...

Data Vortex to introduce their first scalable fine-grained gather-scatter parallel computer at Supercomputing 2014 ...

Kitware works with NVIDIA to drive massive acceleration in HPC visualization ...

Lenovo showcases high-performance computing innovations at Supercomputing 2014 ...

New Orleans becomes home to fastest Internet hub in the world ...

PGI high performance computing compilers coming to IBM POWER systems ...

Mellanox InfiniBand leads TOP500 as most used interconnect technology for high-performance computing for the 2nd consecutive time ...

Mellanox introduces programmable network adapter product line with application acceleration engine ...

Mellanox delivers the world's fastest EDR 100Gb/s InfiniBand switch with latency less than 90 nanosecond ...

Minnesota Supercomputing Institute selects Mellanox EDR 100Gb/s InfiniBand for large-scale supercomputer ...

Rogue Wave reveals new releases and a presentation theater at SC14 ...

ANSYS and NCSA achieve supercomputing milestone ...

Seagate announces reseller agreement with SGI ...