----------------------------------------------------------------------- ; Copyright (C) 1995 ; Associated Universities, Inc. Washington DC, USA. ; ; This program is free software; you can redistribute it and/or ; modify it under the terms of the GNU General Public License as ; published by the Free Software Foundation; either version 2 of ; the License, or (at your option) any later version. ; ; This program is distributed in the hope that it will be useful, ; but WITHOUT ANY WARRANTY; without even the implied warranty of ; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the ; GNU General Public License for more details. ; ; You should have received a copy of the GNU General Public ; License along with this program; if not, write to the Free ; Software Foundation, Inc., 675 Massachusetts Ave, Cambridge, ; MA 02139, USA. ; ; Correspondence concerning AIPS should be addressed as follows: ; Internet email: aipsmail@nrao.edu. ; Postal address: AIPS Project Office ; National Radio Astronomy Observatory ; 520 Edgemont Road ; Charlottesville, VA 22903-2475 USA ----------------------------------------------------------------------- AIPS DDT benchmark results for SUN's SPARCstation 2GX (SUN 4/75) November 28, 1990 Mark Calabretta, Henrietta May Australia Telescope National Facility 1. Introduction --------------- AIPS memo 65 presented DDT benchmark results for the IBM RISC system 6000 model 320 and compared them against a Convex C210. DDT results for several SUN systems were also presented but, as emphasised by the authors, these were not obtained under comparable benchmarking conditions. On 26-27/Nov/1990 we benchmarked a SUN SPARCstation 2GX (SUN 4/75) under conditions suitable for comparison with those of the IBM machine. The basic configuration was as follows 1x 669Mb SUN0669 disk drive for AIPS with a separate drive for unix 1x Exabyte tape drive used for installing AIPS 32Mb memory Graphics accelerator (standard on the 2GX) SunOS 4.1.1_BETA1.0 FORTRAN f77 v1.4 (pre-release) In the first instance we simply copied the complete AIPS directory tree and DDT data files to the benchmark machine from ATNF Epping using tar. The executables, which were completely unoptimized, were compiled under SunOS 4.1, f77 v1.2. After adding the benchmark machine to the list of known AIPS hosts, and enabling the AIPS account and AIPS TV services, we had a fully functioning system, thereby demonstrating compatibility with the SPARCstation 1. This unoptimized system was used to run the small and medium DDT benchmarks. At the request of the SUN representative, we rebuilt the AIPS libraries and DDT executables (INSTEP3) from scratch using the "-fast -O4" optimization level in the f77 v1.4 FORTRAN compiler. This proceded without problems except for $APLNOT/DATBND,DATCAL,GACSIN, and GAININ for which the compiler complained about non-alignment of variables in equivalence statements. These were recompiled without optimization. The compiler also found an invalid integer constant -2147483648 in $APGNOT/FITTP,IMLOD, and UVLOD. However, these tasks are not used in the benchmark, and the problem was ignored. The optimized system was again used to run the small and medium DDTs. The results of the optimized and unoptimized DDT are presented below, together with those of the Convex C210 (selectively optimized) and IBM 320 (optimized) for comparison. The importance of the optimizing compiler for SPARC systems is immediately apparent, as it is for the Convex and, we believe, the IBM RISC systems. AIPS managers will therefore have to confront the problems of the reliability of the optimizer, and the debugging of optimized code. As for the benchmark timings, the ratio of the total real time taken by the SUN to that of the IBM was 1.24 for both the small and large tests. On the other hand, the ratio of CPU times was relatively smaller, being 1.17 for the small DDT, and 1.12 for the medium DDT. This indicates that the SPARC 2 is slightly I/O bound by comparison with the IBM. We believe that SUN will counter this by introducing disk striping in the near future. However, SPARC 2s may be disadvantaged in a networked system where they must rely on an NFS disk server for AIPS data storage. The accuracy measurements quoted in AIPS memo 65 for the C210, IBM 320, and SUN 4/60 (SPARC 1) are reproduced below, together with that of the SPARC 2. There was no significant difference between the optimized and unoptimized versions. Although the results are acceptible, it is interesting to note that the accuracy for the SPARC 2 is systematically less than for the SPARC 1. In subjective terms the SPARC 2 handles beautifully. Not surprisingly, it seems significantly faster in interactive use. Graphics handling, which includes scrolling and AIPS TV display (SSS), benefits noticably from the GX accelerator. The FORTRAN compiler is also much faster, despite the extra overhead of code optimization. 2. Benchmark timings -------------------- SMALL TEST Convex C210 IBM 320 SUN 4/75 SUN 4/75* real CPU real CPU real CPU real CPU UVSRT 11 6.70 14 3.98 19 5.34 16 7.80 UVDIF 7 2.77 10 2.49 6 2.47 6 4.14 CCMRG 4 1.97 2 0.91 2 0.87 2 1.12 SUBIM 3 2.42 2 1.49 2 1.32 3 1.72 COMB 16 10.17 15 6.62 22 6.44 11 7.77 UVMAP 7 5.04 9 4.97 12 6.94 16 12.45 APCLN cln 17 14.25 60 57.97 59 53.36 148 144.87 APCLN res 6 3.36 4 3.05 6 3.62 10 8.36 ASCAL 10 8.05 26 24.54 43 39.67 52 49.63 MX map 11 7.42 8 5.47 14 8.90 23 19.37 MX cln 36 30.82 89 82.54 109 97.13 176 168.01 VTESS 25 16.98 20 15.54 27 18.76 51 45.19 total 153 109.95 259 209.57 321 244.82 514 470.43 ratio 1.00 1.00 1.69 1.91 2.10 2.23 3.36 4.28 MEDIUM TEST Convex C210 IBM 320 SUN 4/75 SUN 4/75* real CPU real CPU real CPU real CPU UVSRT 14 9.31 21 5.87 58 8.91 54 12.55 UVDIF 8 4.38 12 4.12 7 3.83 10 6.90 CCMRG 6 3.10 3 1.60 3 1.65 4 2.44 SUBIM 5 3.84 5 3.95 7 3.10 6 4.38 COMB 22 16.33 38 15.75 77 14.14 73 18.20 UVMAP 14 10.87 26 14.22 45 22.50 64 41.15 APCLN cln 84 77.14 529 517.12 492 455.13 1403 1367.76 APCLN res 14 10.56 17 11.91 40 16.72 53 38.00 ASCAL 20 17.70 77 74.83 268 261.93 322 316.14 MX map 22 17.42 25 17.26 55 30.23 86 65.32 MX cln 124 112.96 611 586.12 625 570.66 1108 1058.31 VTESS 47 39.18 68 48.26 95 63.06 182 145.44 total 380 322.79 1432 1301.01 1772 1451.86 3365 3076.59 ratio 1.00 1.00 3.77 4.03 4.66 4.50 8.86 9.53 Notes: *) The results for UVSRT are the sum of 2 runs, likewise UVDIF 2 runs, SUBIM 2 runs, and COMB 8 runs. *) Real times are measured to the nearest second. *) The results for SUN 4/75* are for the unoptimized FORTRAN compiler. *) The ratios quoted are referred to the Convex. 3. Determination of accuracy ---------------------------- SMALL TEST: Peak Rms C210 IBM 4/60 4/75 C210 IBM 4/60 4/75 UVMAP 13.0 13.1 13.3 12.4 18.8 18.2 18.8 17.9 UVBEAM 10.1 10.2 9.9 8.9 16.2 15.8 16.2 15.3 APCLN 18.8 13.6 14.1 14.1 24.1 16.4 20.7 20.7 APRES 16.9 17.0 17.0 17.0 22.2 22.4 22.4 22.3 MXMAP 12.6 12.3 12.9 11.8 18.5 18.1 18.7 17.9 MXBEAM 13.9 11.1 13.7 11.4 19.3 17.7 19.3 17.6 MXCLN 14.8 14.3 14.3 14.3 17.6 17.6 17.6 17.6 VTESS 4.1 4.1 4.1 4.1 10.9 10.9 10.9 10.8 MEDIUM TEST: Peak Rms C210 IBM 4/60 4/75 C210 IBM 4/60 4/75 UVMAP 13.5 12.9 13.5 12.6 17.8 18.0 18.1 17.9 UVBEAM 14.0 13.0 13.9 12.7 17.8 18.3 18.5 18.3 APCLN 17.3 12.0 11.8 11.7 23.7 14.8 14.7 14.7 APRES 15.4 15.4 15.4 15.3 21.2 21.4 21.3 21.3 MXMAP 13.0 13.3 12.6 13.0 17.8 18.0 18.1 17.9 MXBEAM 14.3 13.0 13.9 12.2 17.8 18.4 18.5 18.2 MXCLN 10.3 10.0 10.3 10.3 14.2 14.2 14.2 14.2 VTESS 3.3 3.3 3.3 3.2 10.8 10.8 10.8 10.8