Horst Simon is the Deputy Director of Lawrence Berkeley National Laboratory. Outside his executive role at the lab, Simon is a highly respected mathematician in the high performance computing (HPC) community. His expertise and research interests are centered on algorithmic development for sparse matrix operations, large-scale eigenvalue problems, and domain decomposition methods.
Within the HPC community, he plays another significant role as the co-editor of the TOP500 project. His TOP500 colleagues rely on his knowledge and connections to verify the existence of systems that vendors claim to have installed and submitted for inclusion in the biannual TOP500 List.
We caught up with him recently to talk about his contribution to the project and its future.
How did you get involved in the TOP500 project?
Simon: I was not officially involved from the very beginning. I attended the Mannheim Supercomputer Conference several times in the early 1990s and became acquainted with Hans Meuer, Erich Strohmaier, and their list of supercomputers. And I provided some comments to them in the early stages.
I think I became actively involved in the TOP500 project after their compilation of the November 1994 list. I remember that there were suddenly a large number of SGI systems on the November list that year, which surprised them. Hans asked if I could check on these newly installed systems, as I was closer to many of the installations, and knew the HPC community in the US well. It turned out that the SGI claims were correct. From then on, I worked with Hans, Erich, and Jack Dongarra each year to do a number of random checks, using my network of colleagues to ensure the systems actually existed.
Convinced of the value of TOP500 for establishing HPC trends, I wrote a paper in 1994 that looked at the claim that Japan was surpassing the US in HPC. My TOP500 analysis showed that this assertion was not correct. Eventually Hans, Jack, and Erich officially invited me to join the team.
What is your present contribution to the project today?
Simon: I continue to review submissions to the list and provide advice to my colleagues. I remember that I argued as early as 2004 or 2005 that we should include energy usage data in the list, but Hans was not convinced.
Had we acted at the time there would have been no need for the Green500. But I do think that the Green500 is very valuable, because it encourages vendors and sites to think of energy savings as an important goal.
My long involvement with the list and the data has helped me to quickly hone in on trends and help publicize them. In 2012 I became convinced for the first time that HPC performance is actually slowing down. I made this point subsequently in many presentations. I also used the Gini coefficient to establish that we have experienced a trend toward much larger systems, which has made the list more top-heavy.
It has been more than 20 years since the TOP500 project started. What transformations has it experienced and how has it transformed the nature of high performance computing?
Simon: Our team just wrote a paper on that topic that will appear in the November 2015 issue of IEEE Computer: “The TOP500 List of Supercomputers and Progress in High Performance Computing.” This paper addresses the question you pose in much more detail. Actually the list itself, the use of High-Performance Linpack (HPL), and how the ranking is done, has stayed essentially the same for this period of time. This is in itself remarkable in a field of exponential growth and rapidly changing technology.
I am coming to realize that the greatest value of the list is that its methodology did not change and thus it delivers the ability to compare systems and trends over a long time span, from the giga to the peta era, which is a factor of a million!
In spite of frequent criticisms, I believe that the list continues to have an overall positive influence on the field. It allows the HPC community to measure progress and document performance growth. Few other fields have such a metric.
The architecture of supercomputers has evolved - homogenous vector machines to heterogeneous machines based on commodity components - and so has the usage. In particular, there’s more industrial usage of HPC than ever before. How is the TOP500 project relevant to this newer community?
Simon: I sincerely hope that industrial users will not use the TOP500 list to decide on what products to buy. They should rather use the list to see what others in their field are doing and what their competitors are investing in HPC. A while ago I saw a presentation on investments in HPC and productivity in the automotive sector. It was amazing to see how well the two correlated.
What does the future hold for the TOP500 project?
Simon: I expect to be active in the field for at least another decade, so I would hope that the TOP500 list lasts at least that long.
The middle of next decade will be an important time for HPC. I do expect that we will have exascale systems by then, and that we will continue to rank them by HPL and the TOP500 list.
But as I said, I also see performance slowing down, and it will be difficult to foresee what will happen by 2030. Computing technology will become more diverse and we may use very different architectures, such as neuromorphic computers. At that point, the HPC community will need to rethink performance evaluation and how to compare such diverse systems.