Cross-town athletic rivals UCLA and USC are playing out a new race with UCLA’s debut at position 148 on the Top500 list with a 68 TFlop HP ProLiant SL390s system. Comparing and contrasting UCLA’s new entry with USC’s position 95 94 TF system provides some fascinating insight to the rapid advances in supercomputing technology as well as the history of supercomputing in Los Angeles. Lets take a look, starting with the USC system.
USC definitely scores high points for longevity. USC has continuously placed an x86 system on the Top500 list since June 2003 when they just broke through the TF boundary with a 1.028 TF, #55 ranked system. For 17 Top500 lists, USC has rebuilt and expanded this cluster, ranking as high as #26 on the June 2006 list with a 13.8 TF system. I haven’t checked the history of every Top500 cluster, but USC’s entry in 17 consecutive Top500 lists has got to be some sort of record. USC kept their Top500 entry unchanged from November 2010 to this week’s list and thus the same 94 TFs dropped USC from November’s position 80 to today’s position 95. USC’s approach of continually refreshing their cluster is evident in their Top500 entry which lists the system as containing Sun, Dell, and IBM nodes with both Intel and AMD x86 processors. USC is also one of the few sites that still uses a Myrinet interconnect, not surprising given that the original Myrinet technology was developed in 1991 by a research group at USC’s Information Sciences Institute. According to USC’s HPC site, their cluster does contain a number of newer HP ProLiant SL160 compute nodes, although it appears that USC did not have time to re-run their Top500 submission since installing the HP nodes. That happens sometimes when you have real researchers using the system fulltime for scientific discovery (vs benchmarking).
Now lets drive across town and take a look at the UCLA system. At 68 TF, the UCLA system has about 30% less performance than USC’s. But that isn’t the most interesting part. First of all, a huge congratulations to UCLA for their first Top500 entry since their June 2006 entry which barely made the Top500 cutoff at position 461. For five years, if you went to school in Westwood or were a researcher there, the closest Top500 supercomputer facility was across town at USC. Historical note, the 2006 system at UCLA was based on Apple Xserve servers. No one needs to ask why that system didn’t stay on the Top500 for long.
Fast forward to today. UCLA’s latest system is based on HP’s ProLiant SL390s G7 servers, with 2482 Xeon processor cores. That compares with USC’s 13000 Xeon and Opteron processor cores. Wait a second, you might say, is that a mistake? Why does USC’s system, with 5x as many x86 processor cores, only provide 30% more performance? The secret, of course, is the Nvidia M2070 GPUs integrated into the SL390s. Like Tokyo Tech’s #5 ranked system and Georgia Tech’s #169 ranked system, the UCLA system achieves amazing FLOPs/watt/$ by using Nvidia GPUs. Sure, some people have claimed that using Nvidia GPUs to place high on the Top500 system is “gaming” the system, heck, some people claim the entire Top500 list is irrelevant. But the reality is that across the world, including on 3 of the top 5 fastest supercomputers in the world – which happen to be far across the Pacific in Japan and China, to right in my hometown at UCLA, researchers are increasingly turning to hybrid systems including GPUs from Nvidia, from AMD, and in the future Intel’s MIC, to achieve amazing scientific results, using systems costing far less to deploy, and using only a fraction of the energy of pure x86 clusters.
You can ask, but I won’t comment in this public forum what team I’ll be cheering for when I go to the UCLA-USC football game this November, but you can bet that when the November Top500 list is unveiled at SC11 that same month, I’ll be cheering for both UCLA and USC to increase their rankings!