Recently, I heard David Keyes talk about exascale computing ("Exaflop/s Seriously!"). There were two interesting lessons.
1. In one of his viewgraphs (which I unfortunately do not have access to) he plotted the peak performance of the top supercomputer in the Top500, versus time. The resulting curve looks familiar (Moore's law), and when drawn on a log-linear scale looks approximately linear. I reproduce a similar looking curve (but this one combines all the 500 top machines) below.
The interesting part of the viewgraph was that, on the same plot, he also threw in the performance of the 500th best machine in the Top500. This was also linearly increasing (on a log linear plot) curve that lay somewhat lower than the fastest computer. The surprising bit (for me at least) was that the "phase shift" in terms of time was about 8 years. That is, if the top machine in the Top500 was left alone, in 8 years it would be overtaken by everybody else in the elite group.
This can have policy implications for HPC centers. Rather than trying to build the fastest computer now, perhaps, having a long-term plan to keep upgrading the system is more important.
2. We've seen CPU clock cycles stagnate. For HPC centers, which use thousand of them, energy considerations can quickly become the dominant concern. This explains the shift to multicore. I learned that energy scales roughly as the third power of the frequency. From here, for example
I remember that we had a small Beowulf cluster (puny by today's standards), when I was in grad school, that was housed in a student office for sometime. The room was unbearable in summer.
1. In one of his viewgraphs (which I unfortunately do not have access to) he plotted the peak performance of the top supercomputer in the Top500, versus time. The resulting curve looks familiar (Moore's law), and when drawn on a log-linear scale looks approximately linear. I reproduce a similar looking curve (but this one combines all the 500 top machines) below.
The interesting part of the viewgraph was that, on the same plot, he also threw in the performance of the 500th best machine in the Top500. This was also linearly increasing (on a log linear plot) curve that lay somewhat lower than the fastest computer. The surprising bit (for me at least) was that the "phase shift" in terms of time was about 8 years. That is, if the top machine in the Top500 was left alone, in 8 years it would be overtaken by everybody else in the elite group.
This can have policy implications for HPC centers. Rather than trying to build the fastest computer now, perhaps, having a long-term plan to keep upgrading the system is more important.
2. We've seen CPU clock cycles stagnate. For HPC centers, which use thousand of them, energy considerations can quickly become the dominant concern. This explains the shift to multicore. I learned that energy scales roughly as the third power of the frequency. From here, for example
"If the clock rate of the Multi-Core CPUs will be reduced by 20% only, then the energy consumption is reduced to 50% compared to a system running at full clock speed.”On the other hand, for multicore CPUs the energy is only proportional to the first power of the number of cores, explaining the move towards multicore, low frequency supercomputers. I don't know if the numbers here include the cost of air-conditioning (I presume they do), but they can be a significant operating cost.
I remember that we had a small Beowulf cluster (puny by today's standards), when I was in grad school, that was housed in a student office for sometime. The room was unbearable in summer.
No comments:
Post a Comment