From: gaume@nh3.usno.navy.mil (Ralph Gaume) Subject: AIPS DDT results for IBM RS6000 3BT at USNO Date: Tue, 20 Dec 1994 11:21:00 -0500 Pat and Eric- I thought you may be interested in the results of my recent port of 15JUL94 AIPS to our 2 new IBMs here at the Naval Observatory. As far as I know, this is the first port of AIPS to IBM's Power2 series cpu and new Fortran90 compatible compiler. With an AIPSmark(93) of 4.6, these machines are the fastest AIPS system reported to date. Below I have listed several notes on the port and the DDT test, along with data from the DDT tests to fill in the 9 tables of Eric's AIPS Memo (#85) on DDT measurements of various platforms. -Cheers, Ralph NOTES ON PORT AND DDT TESTS 1) Fortran compiler = IBM's xlf version 3.2.0.0 C compiler = IBM's xlc version 1.3.0 OS = AIX version 3.2.5 The fortran preprocessors (-Pv and -Pk options) no longer come with the compiler for free, and were not used in this port. The fortran compiler has options for optimizing for the Power2 architecture. These options were not applied in this port of AIPS. 2) FITTP experienced an error writing UV data files to FITSdisk or FITStape. FITTP worked correctly with MA data files. FITTP was found to work properly with UV data files by recompiling and omitting the -O fortran compiler option. 3) I experienced significant problems reproducing total elapsed times for various repetitions of the DDT. I traced this to unusual problems with name resolution in the USNO nameserver (the AIPS MSGSRV makes extensive use of the nameserver). The name resolution problem was reported to USNO network personnel, and the IBM taken off the network (with internal name resolution) for subsequent executions of the DDT. These DDTs were then reproducible to 0.6% in total elapsed time. 4) To maximize I/O throughput, DDT I/O was spread across 4 external SCSI-2 F/W DIFF Barracuda 4 (4Gby) disks (all on the same SCSI controller). 5) Hardware configuration: (standard features) IBM RS6000 model 3BT (called a 390 in server configuration) Power 2 chip @ 67MHz Integrated SCSI-2 F/W SE dual ported controller 32 kby instruction cache 64 kby data cache Hardware configuration: (upgrades/options purchased) (1) internal 2Gby F/W SE IBM disk (upgrade) (1) external SCSI-2 F/W DIFF controller (option) (4) Seagate Barracuda 4 (4Gby) SCSI-2 F/W DIFF disks (external) (option purchased from NPI) (1) 128 Mby memory (upgrade) (1) 1Mby L2 cache (option) (1) Gt4xi 24-bit graphics card (upgrade) (1) internal CD-ROM-2 (option) (1) 4.0 Gby 4mm DDS-2 tape drive (option) (1) 5.0 Gby 8mm tape drive (purchased from TTi) (1) 20" monitor (upgrade) DATA TABLES FOR AIPS MEMO 85 Table 1 ------------------------------------------------------------------------ | Computer Name OS OS version Memory | | IBM RS6000 (3BT) quasar AIX 3.2.5 128 | ------------------------------------------------------------------------ Table 2 ------------------------------------------------------------------------------- | Computer compiler level 0 control code OPT2 adds OPT6adds | | IBM RS6000 (3BT) xlf -c -u -qfips -qcharlen=10000 -O -Q -ND16384 nothing | | -qextname -NA65536 | | -qmaxmem=32768 | ------------------------------------------------------------------------------- Table 3 This is a new machine, no old DDT results are available Table 4 Computer Am(93) TLarge TApcln TCalib TMxcln TVtess Trest IBM RS6000 (3BT) 4.6 869 254 62 301 57 195 Table 5 Computer APCLN CALIB MXCLN VTESS ratio cpu ratio cpu ratio cpu ratio cpu IBM RS6000 (3BT) 0.94 240 0.53 33 0.89 268 0.68 39 Table 6 Computer TLarge TMedium TSmall IBM RS6000 (3BT) 869 256 137 Table 7 Program max rms UVMAP 12.3 15.7 UVBEAM 13.7 16.3 APCLN 9.8 16.5 APRES 16.3 23.2 MXMAP 13.9 15.2 MXBEAM 14.4 15.4 MXCLN 10.2 16.5 VTESS 21.7 30.0 Table 8 IBM RS6000 (3BT) MFLOPS= 55.1 Linpack DP 73.0 Linpack SP 182.8 Linpack TPP SPECint92 = 114.3 SPECfp92 = 205.3 R = 9.8 Am(93) = 4.6 ALm(93) = 4.8 Table 9 Computer Loop 1 Loop 2 Loop 3 Loop 4 Loop 5 Loop 6 Total Real ALm(93) IBM RS6000 (3BT) 3.0 38.0 21.0 3.0 27.0 34.0 125 126 4.8