The keynote speech given by Dr. Jack Dongarra at HPC Asia 2009 examines the history of high performance computing (HPC) from its beginnings in the 1950s, through the present, and into the very near future. The talk also takes an in-depth look at the TOP500 supercomputing list which was begun by Dr. Dongarra and several of his colleagues in 1993.
This extremely educational and enlightening talk also takes a look at current trends in HPC such as “many-core” chips and GPUs as well as examines future obstacles in the ongoing development of HPC.
Part1 | Part2 | Part3 | Part4 | Part5 | Part6 | Part7 [PDF Download] [Video]
Part 3ˇGTOP500: Processors, Interconnects, and Efficiency
If we take a look at what kinds of processors the machines on the TOP500 list are using, we can see that Intel is used in 71% of those machines, AMD is used in 13%, IBM has about 7%, and there’s a few leftover. So Intel clearly dominates in terms of what’s inside of these high performance machines. Next, let’s take a look at the interconnects that are being used in these machines. The latest slice of data from November 2008 says that about 270 machines are using GigE as the interconnect. GigE is a rather inexpensive, rather pour interconnect to be using in a supercomputer, but it is used in so many of them because of its relatively low cost. What’s happening here is that vendors are low-balling the price of their machines, putting together lots of processors, and then connecting them with a rather poor interconnect. Anyway, moving up, InfiniBand is used in about 140 machines and then we have Myricom and Quadrex which are used in just a few of them. The most striking thing here though is that GigE is used in so many of these systems today.
Now let’s take a look at the efficiency of the 500 machines on the list. Think of the efficiency as the ratio of achieved to peak performance of your computer. To do this, all you need to do is figure out how many operations you can do per cycle and look how many processors and cores you have in your machine. That’s your theoretical performance. And then look at the actual performance that you obtain running the benchmark and the ratio of that is your efficiency. We see that some of these machines are as high as 90% efficient running the benchmark. We see a big range as well--from 90% all the way down to 30%. We also see this line here between 50% and 60% efficiency. That’s the line where the machines that use GigE are. GigE, a rather poor interconnect, that shows up quite clearly in terms of the performance that we see from these machines. This machine at 30% efficiency, I can guarantee you, they’ve made a mistake in putting together the benchmark. They’ve done something wrong. It should be much greater than 30% even with a poor interconnect. That’s clear.
Today the majority of sockets on the Top500 list have four cores. There a few machines that have single core sockets. There’s a grouping of machines, about 30%, which have dual core sockets, and then the remaining 2/3 have quad core sockets. Then there are seven systems which have a chip that uses nine cores--the IBM Cell.