Dec 14, 2008

The iPAB open seminar "Trend and future of HPC"



The open seminar of the initiative Parallel Bioinformatics(iPAB) was held on December 5, featured "Trend and future of HPC". This open seminar was focused on high performance computing (HPC) indispensable to bioinformatics.

Besides iPAB, Chem-Bio Informatics Society (commonly called CBI Gakkai) with a long history, and Japanese society for bioinformatics (JSBI) that held the annual meeting in Senri on 15th - 16th of December are known as an academic active group in Japan which intends to involve HPC for the life science.

They have their own characteristics, and in my impression, iPAB seems to have the keen interest in an immediate application of HPC, JSBI is active in education and basic research of bioinformatics, and CBI Gakkai works in broad area relative to the drug discovery.

Well, three following HPC lectures were literally done with the theme in "Trend and future of HPC" open seminar in the O-okayama campus of Tokyo Institute of Technology by iPAB. Because of that, it was few attendees from the life science area, and many of participants are related to HPC.

Lecture 1. "SC08 report" by Prof. Yoshio Oyanagi, Kogakuin University

Prof Oyanagi's SC08 annual report are revised on 12/7.

Lecture 2. "Trend and future of HPC" by Prof. Taisuke Boku, Tsukuba University

・ It was mainly a lecture on technology trends of the cluster system. It was a good and comprehensible lecture taking T2K-Tsukuba large-scale, general-purpose cluster for example (T2K means Tsukuba, Tokyo and Kyoto).

・ It is completely agreed, in particular, for the point that the gap between the floating point arithmetic performance and the memory bandwidth is extending more and more because the processor with an extremely high floating point arithmetic performance has come to be used with a commodity HPC cluster system and it becomes a serious issue. For observation that many-core chip could not significantly contribute to HPC as a result, Prof. Taisuke Boku suggested that surplus cores might have to be used for the purposes other than the floating point calculation.

・ Moreover, about accelerator of much interest in SC08, he said that because accelerator is comparatively easy to implement on the cluster system , the aggressive use can be just conceivable. In that case, the data transfer bandwidth to accelerator must be the severest issue. Also the issue of the reliability of GPGPU (e.g., There is no ECC in a built-in memory of the GPU board) was discussed.

Lecture 3. "The latest TSUBAME system", Prof. Satoshi Matsuoka, Tokyo Institute of Technology

・ It was a lecture about the TSUBAME 1.2 system of Tokyo Institute of Technology. TSUBAME is known well as a system that has positively taken accelerator at first and now 648 ClearSpeed accelerators are being attached with it according to Prof. Matsuoka. In addition, the performance was improved by enhanced 318 nodes added Tesla S1070 accelerator. Then TSUBAME won the second place in Japan succeeded to the T2K of University of Tokyo in the Top500 list of November, 2008.

・ Thus the node of TSUBAME 1.2 has three types: (A) Opteron + ClearSpeed + Tesla (318 nodes), (B) Opteron + ClearSpeed (330 nodes), and (C) Xeon (90 nodes). It is a really heterogeneous commodity cluster system. It may be beyond expectation in general that such a heterogeneous system realizes better LINPACK performance in short period after the version up, and typical cluster users could never do something like that.

・ According to Prof. Matsuoka, GPU resembles a classical vector computer because of the feature of the high peak performance and a high memory and the bandwidth (in GPU). However he also pointed out that GPU also has some technical limitations and further study is required as shown in the history of PC cluster becoming main stream of a general-purpose parallel computer. Such situation imagines me the enough possibility that some Japanese computer manufacturer might develop their original accelerator in the future.

TSUBAME 2.0 is supposed to be run in the spring of 2010. Do not miss it in one and a half years.


Although I did not know details of TSUBAME, the previous talk by Prof. Bok helped me to spend interesting time to understand TSUBAME much well. Well, except for those having high level skills and experiences enough to achieve implementation like Tokyo Institute of Technology, it looks that the entire cost will never become cheap of the expectation from the word commodity cluster system for common users excluding particular universities. In addition, since application porting/tuning efforts are mandatory for such a peta-scale cluster systems to run applications with reasonably high performance, the word challenge is more suitable than exploiting. Grand Challenge is just exact word.

No comments:

Post a Comment