krowitz@EDDIE.MIT.EDU@mit-kermit.UUCP (02/18/87)
I just installed SR9.5 on our network and began recompiling our programs to take advantage of the new compilers which (it is claimed) produce code that is significantly faster than the old SR9.2 compilers produced. Just for fun I tried two of my old favorites, the Whetstone and Linpack benchmarks. Here are the results: Whetstone (single) Whetstone (double) ------------------ ------------------ 330 (SR9.2) 682 665 330 (SR9.5) 744 668 560 (SR9.2) 794 775 560 (SR9.5) 848 761 660 (SR9.2) 757 471 660 (SR9.5) 678 436 As you can see, the DN330 and DN560 got about 10% faster in single precision, but had no increase in double precision. Even more curiously, the DN660 got 10% slower in both single and double precision. This is a performance enhancement? Why would single precision get better and not double precision? The story gets even worse when you look at the Linpack results: Linpack (single) Linpack (double) ------------------ ------------------ 330 (SR9.2) .072 .063 330 (SR9.5) .068 .060 560 (SR9.2) .079 .073 560 (SR9.5) .075 .068 660 (SR9.2) .102 .067 660 (SR9.5) .098 .061 In the case of Linpack, all nodes get 5% to 10% slower with the new compilers. What is going on here? Have any other sites using SR9.5 (or beta-testing it) found a real performance gain? -- David Krowitz mit-erl!mit-kermit!krowitz@eddie.mit.edu mit-erl!mit-kermit!krowitz@mit-eddie.arpa david@mit-mc.arpa (in order of decreasing preference)