in the November 2003

16 November 2003

Two Myrinet Clusters in the Top 10

The TOP500 list (www.top500.org), published twice a year prior to the ISC conference in June and the SC conference in November, ranks supercomputers worldwide according to their performance on the LINPACK benchmark. The 22nd TOP500 list was published on 16 November 2003 prior to SC2003 in Phoenix.

Myricom's Myrinet® cluster-interconnect technology made additional gains in the November 2003 TOP500 list, but the most striking news is the two Myrinet clusters in the top 10.

• The Tungsten cluster at the National Center for Supercomputing Applications (NCSA) is the first Myrinet cluster to make it into the top 5. Tungsten ranked #4 at 9819 Gflops, in spite of being not fully operational. The High-Performance LINPACK (HPL) benchmark was run across 1,250 of Tungsten's 1,280 Dell PowerEdge 1750 dual 3.06GHz Pentium-4 nodes. With the full complement of nodes and some software tuning, we expect Tungsten to reach ~11000 Gflops for the June 2004 TOP500. Tungsten was supplied to NCSA by Dell, and is the highest ranking Dell cluster and Intel cluster in the list.
• The Lightning cluster at Los Alamos National Laboratory (LANL) ranked #6 at 8051 Gflops. Lightning is composed of 1,408 dual 2GHz Opterons, and was supplied to LANL by Linux Networx. The Lightning HPL result of 8051 Gflops is 71.5% of peak, which is excellent. The November 2003 TOP500 is the first to include systems based on the AMD Opteron processor. Lightning is also unusual in being a LinuxBIOS/BProc cluster; its nodes boot entirely over the Myrinet.

The November 2003 TOP500 list includes a total of 73 Beowulf-style Myrinet clusters, up from 51 in the June 2003 list, and up from 28 in the November 2002 list (only one year ago!). Of these 73 Myrinet clusters, 34 are new or upgraded, and 39 carried over from the June 2003 list. Eleven of the 73 Myrinet clusters are listed as self-made.

In addition, the November 2003 TOP500 list list includes 121 HP Superdome^TM/Hyperplex^TM systems, down from 127 in the June 2003 list. HyperFabric^TM is an HP brand for Myrinet communication between HP Superdome SMP hosts. We understand that most of the higher performance Hyperplex systems use HyperFabric. Based on performance and processor counts, we believe that all of the HP Superdome/Hyperplex systems in the November 2003 TOP500 list use HyperFabric, hence Myrinet technology.

Thus, a total of 194 (38.8%) of the November 2003 TOP500 supercomputer sites use Myrinet technology.

Here is our tabulation of the 73 Myrinet clusters in the November 2003 TOP500. Most of the Myrinet clusters are identified in the list as such, and others (e.g., #189) are known to us to be Myrinet clusters. Similarly, if the computer has a name, we list it below, even though that name may not appear in the TOP500 list. The full configurations of these clusters, including the types and numbers of hosts, can be found in the list at http://www.top500.org/.

Congratulations to the TOP500 authors, and to all of the TOP500 sites. We are looking forward to the June 2004 TOP500 list.

16 November 2003