Figures 1 and 2 display select results of a benchmark comparision between the new NPAC compiler the PGI HPF compiler, version 2.0. The programs were run on the IBM SP2 at NPAC.
The Laplace benchmark performs Jacobi relaxation on 256*256, 512*512 and 1024*1024 arrays, distributed blockwise over various numbers of processors. Both compilers achieve about the same perfomance on a single node, but generally our compiler exhibits better speedup on multiple processors, presumably due to more effective handling of communication.(See Figure 1)
The synthetic benchmark involves no communicaion -- it is a forall assignment involving large arrays. It illustrates that (unlike the PGI compiler) we deal with address translation efficiently, even for cyclic destribution format.(Speedup is relative to an equivalent sequential program compiled with the IBM Fortran compiler.) (See Figure 2)
While these examples are necessarily select, in genetal we find that (on code that both compiler can successfully compile) the NPAC compiler compares very favourably with the commercial compiler.
Back @