SCD moves to highly parallel, DSM computers

Users encouraged to become familiar with this technology . . .

Bill Buzbee Contents Search Article index Back issues Subscribe Contact us SCD

by Bill Buzbee As noted in the director's column of the last issue of SCDzine, NCAR is well underway with plans to transition to highly parallel computers with Distributed Shared Memory (DSM) microprocessor systems. The first DSM (sioux) was installed last summer and is now available to community users. There is no General Accounting Unit (GAU) charge for using this machine, and anyone with a GAU allocation may request an account on sioux. This month (May 1998), a 128-processor SGI Origin 2000 system (ute) will be installed in the Climate Simulation Laboratory. Following are measured performances of a couple of NCAR models, as well as models from other labs and from the University of Miami on the SGI Origin. NCAR and other models Sustained megaflops/processor on SGI Origin2000, 195MHz   No. of processors Institution Model Model type 8 16 32 64 128 NCAR CCM3.2 Atm. GCM 78.5 71.8 61.4 56.0   NCAR/PSU MM5 Mesoscale   111.1 98.4 73.5 48.4 LANL POP Ocean GCM 51.6 54.6 53.0 42.4   U. Miami SC-MICOM Ocean     85.3 74.6 59.8 U. Miami MP-MICOM Ocean 92.3 94.6 93.7 93.6   NASA Ames CFL3DHP CFD 35.8 35.4       NASA Ames TLNS3DMP CFD   50.6 39.2     NASA Ames Overflow CFD       51.5   NASA Ames ARC3D CFD       98.4   Average 64.6 69.7 71.8 70.0 54.1 Note that the metric is performance per processor. Thus, when a model uses 64 processors, we believe that this machine will be approximately equal to the C90 in capability, i.e., about 5 gigaflops. Budget permitting, we expect to install additional DSMs in FY99. So we strongly encourage you to become familiar with this technology.

Contents || Search || Article index || Back issues || Subscribe || Contact us || SCD