================================================================== === === === GENESIS / PARKBENCH Parallel Benchmarks === === === === POLY1 === === === === R-infhat and F-half === === In-Cache Memory Bottleneck === === === === Versions: Std F77 === === === === Author : Roger Hockney === === Department of Electronics and Computer Science === === University of Southampton === === Southampton SO9 5NH, U.K. === === fax.:+44-703-593045 e-mail:rwh@uk.ac.soton.pac === === vsg@uk.ac.soton.ecs === === === === Last update: November 1993; Release: 1.0 === === === ================================================================== 1. Description -------------- This benchmark tests severity of memory bottlenecks by varying the amount of arithmetic per memory reference which is called the computational intensity of the loop. The performance for long loop (vector) lengths, RINF, is represented as : RINF = RHAT/(1 + FHALF/F) (1) where RHAT = peak Mflop/s rate of arithmetic pipeline approached as F goes to infinity and F = computational intensity = ratio floating operations/memory references FHALF = F required to obtain RINF=RHAT/2 The loop executed is polynomial evaluation by Horners rule, where the computational intensity is equal to the order of the polynomial. The order and F is increases from 1 to 10, and the results for RINF for each value of F are fitted by least squares to equation (1), giving the best value of the parameters RHAT (R-infinity-hat) and FHALF (half-performance intensity) for this fit. POLY1 chooses vector lengths that fit into the cache, and FHALF is a measure of the ratio arithmetic performance (Mflop/s) to cache-memory access rate (Mword/s). For further details of the FHALF characterisation, Hockney and Jesshope, Parallel Computers-2, IOP Publishing, Bristol and New York, Chapter-1. 2. Operating Instructions ------------------------- To compile and link the benchmark type: `make' . On some systems it may be necessary to allocate the appropriate resources before running the benchmark, eg. on the iPSC/860 to reserve a single processor, type: getcube -t1. To run the benchmark type: poly1 Output from the benchmark is written to the file "poly1.res" If the timing results are too inaccurate the parameter NITER in file poly1.inc may be increased. This is the number of repetitions of the kernel loop used to extend the length of time measured. NITER=1000 is a sensible starting value. NITER=10 may be used for testing execution but is probably too small for accurate timing. The order of executing of the kernel loop should be as specified in the Fortran code (in SUBROUTINE DOALL). Nonesense results (e.g. negative FHALF) may be produced if the compiler tampers with the loop ordering or does software pipelining. The polynomial must be completely evaluated for one value of the loop index-I (e.g. DO 310 loop) before the next value of I is taken. evaluated. $Id: ReadMe,v 1.2 1994/05/25 16:54:25 igl Exp igl $

Submitted by Mark Papiani,

last updated on 10 Jan 1995.