name: Ron Sercely establishment: Convex Computer Corporation email: sercely@convex.com telephone: 214.497.4667 date sent: 29/Nov/95 date benchmarked (DD/MMM/YY): 29/Nov/95 benchmark suite and revision: Parkbench message passing library and revision: PVM 3.3.10.2 (to be released Dec '95) benchmark: comms3 problem size (if applicable): number of nodes: 4 file number if one of a group: 3 of 7 manufacturer: Convex model number: Exemplar 1200 cpu: HP-PA/RISC 7200 cpu speed (MHz): 120Mhz fpu: integrated primary cache: 250kbytes secondary cache: none other cache: none memory: 512M disk and connection: SCSI interconnect type: shared memory switching: none other hardware: none operating system and version: SPP-UX 3.2 (to be released Dec ' 95) compilers, libraries and versions: Convex FORTRAN, version 9.3 Convex CC, version 6.3 compiler switches: -O2 (full optimization) additional tuning parameters: other software: additional comments: PARKBENCH Version 1.0 *** UNOFFICIAL release test package *** This run was with modified code. Modifications were: Within the timing loops, the code segments like: CALL PVMFINITSEND(PVMDEFAULT,INFO) CALL PVMFPACK(BYTE1,A,ILEN,1,INFO) CALL PVMFSEND(TIDS(NSLAVE),10,INFO) CALL PVMFRECV(TIDS(NSLAVE),20,INFO) CALL PVMFUNPACK(BYTE1,A,ILEN,1,INFO) was changed to: CALL PVMFPSEND(TIDS(NSLAVE),10,a,ilen,byte1,INFO) CALL PVMFPRECV(TIDS(NSLAVE),20,a,ilen,byte1,iatid, iatag,ialen,INFO) The PSEND/PRECV functions are Convex versions of SEND/RECV optimized shared memory PVM implementations. It is a documented, standard Convex feature. Also, in synch1, one line of code was inserted just before the timing loop:: CALL PVMFFREEZEGROUP(GROUP,NNODE,INFO) This function is a documented, standard Convex feature, which allows optimized shared memory group operations, such as PVMFBARRIER Environment Variables: PVM_ARCH = CSPP PVMBUFSIZE=1000000 PVMPSEND=0 (enables the PSEND functions) ================================================= === === === GENESIS / PARKBENCH Distributed Memory === === Benchmarks === === === === COMMS3 === === === === Program: Saturation Bandwidth === === Version: PVM + Fortran 77 === === Author: Roger W. Hockney === === === ================================================= Case LENGTH(B) TIME(sec) RAV(B/s) 1 200 6.165E-05 3.244E+06 2 400 2.556E-05 1.565E+07 3 600 2.537E-05 2.365E+07 4 800 2.742E-05 2.918E+07 5 1000 3.019E-05 3.313E+07 6 2000 3.951E-05 5.062E+07 7 5000 5.684E-04 8.797E+06 8 10000 9.961E-04 1.004E+07 9 20000 1.018E-03 1.964E+07 10 50000 1.139E-03 4.392E+07 11 100000 1.351E-03 7.401E+07 12 200000 2.479E-03 8.067E+07 13 500000 8.410E-03 5.945E+07 14 1000000 1.682E-02 5.947E+07 ---------------------------- COMMS3: Saturation Bandwidth ---------------------------- Result Summary -------------- Number of Processors(nodes) in test = 4 Maximum Observed Total Bandwidth = 8.067E+07 Byte/s Maximum Bandwidth per Processor = 2.017E+07 Byte/s