Recent DDT Benchmark Results

Silicon Graphics

1999-07-02: Torsten Ensslin at the University of Toronto has run the benchmark on a SGI Origin 2000 system with 4 300-MHz R12000 processors and 16 MBytes of cache. Main memory size is 512 Mbytes, and disk space was 72 GB with two 18GB user disks striped into a 35 GB XLV (with 1GB log) for super I/O performance. OS was Iris 6.4 and the binaries were the pre-compiled versions from NRAO's Origin 200.

Jeff Pedelty reports an AIPSMark of 14.2 on a 4-CPU Onyx (R10000) SGI system, on a non-empty system. Watch this space for possibly better results on a quiet system...

You can read all about it in AIPS Memo 94, but the gist of it is this: an AIPSMark of 13.7. This was on a SGI Origin200 system with 4 processors. The numbers on the 4 simultaneous DDT's run together is fascinating too. A more recent run of a single DDT on 15OCT98 yielded an AIPSMark of 18.7 when the speed parameter was set to 18.

Jeff Pedelty at Goddard Space Flight Center ported the 15JAN94 AIPS to a Silicon Graphics system in early 1994. Here are several of his AIPSMark(93) figures from different machines (all using Irix 5.2 unless otherwise stated):

SGI Indigo I (33MHz R3000, 16MB, 32KB caches): 1.18
SGI 150MHz MIPS R4400, 32 Megabytes, SCSI-I: 2.66
Same system, 64 Megabytes, Seagate 9G fast SCSI-2 (not wide): 3.47
SGI Power Challenge, single 75Mhz MIPS R8000 CPU, 16KB instruction and 16KB integer data cache, 4MB secondary cache for floating point data, 256 Megabytes, APsize 50 MBytes (Irix 6.0.1): 6.8 (Feb 13).
Same system, used all but one of the Convex $QVEX Q routines (library for FFT's etc): 8.7 (Feb 16).
NEW: SGI 15MHz MIPS R4400 as above, with the -mips2 option for R4000-specific instructions (better optimization at the cost of incompatibility at the binary level for older SGI systems): 4.55

The Power Challenge machine used is part of a 4-CPU Symmetric Multiprocessor system, but only one CPU was active at the time of the test. The R8000 is apparently also available in a desktop package (Power Indigo2) but with a smaller F.P. cache; no information is available on its AIPSMark performance yet.

Thanks to Jeff (pedelty@jansky.gsfc.nasa.gov) for making the information available (and for helping to provide the 15JAN95 and 15JUL95 binary distribution for SGI!).

Patrick P. Murphy