eugene@ames.UUCP (Eugene Miya) (11/05/84)
Recently, a paper in Computer [IEEE], Oct. 1984 mentioned the content of the infamous LLNL "Livermore Loops." I have been aware of the content of these benchmarks for about 6 or 7 years now. Basically, they were written to test "cute" ideas on CDC 7600 class serial processors. [My favorite is the many way cascading branch loop which is supposed to fool lookahead pipelines.] These programs have several problems with future high speed processors: they are written in FORTRAN, they cannot distinguish vector from scalar machines, they cannot distinguish various levels of parallelism (fine grained dataflow from say Denelcor or Cray task-level parallelism). A friend doing dataflow research at LLNL and I are interested in writing a set of test program generators (in somewhat of a `reasonably' language-independent way) as a follow on to the Livermore loops. I choose not to post this to language group because determining performance characteristics (not just speed) is my major concern. The orientation is numerical. We want to be able to detect architectural differences such as dataflow architectures from vector architectures, and so on. If a machine has N CPUs, we want to be able to detect overheads and fall offs. If you are interested in discussing the issues [we don't want to write code without giving this good thought], write or phone me. --eugene miya NASA Ames Res. Ctr. {hplabs,ihnp4,dual,hao,vortex}!ames!aurora!eugene emiya@ames-vmsb.ARPA New phone prefix: (415)-694-6453 (5197 mesgs)
eugene@ames.UUCP (Eugene Miya) (11/09/84)
I am aware that benchmarks in the past tended to give single numbers which ignorant managers would go running off into the void to argue for some stupid machine. The intent of my solicitation for others interested in benchmarking is to come up with a set of tests (don't call them benchmarks then) which would be able to distinguish architectural characteristics of various architectures. Several of us want to be able to determine the extent of pipelining (hardware), dataflow architectures, vector architectures, and so on. I realize this is not an easy task, but I want to give it a try and I have a couple of other researchers also willing. --eugene miya NASA Ames Res. Ctr. {hplabs,ihnp4,dual,hao,vortex}!ames!aurora!eugene emiya@ames-vmsb.ARPA