1. The cc-NUh4A memory architecture of the SGI Origin 2000 can significantly alter the performance of a shared memory implementation depeuding on how the requested memory is mapped to the architecture, since this memory is logically shared, but physically distributed. The current implementation makes use of the first tech rules, in which memory is allocated to the processors which are the first to access or touch it. Memory placement is thus achieved by executing a parallel loop in which each processor initializes all arrays on the subdomain to which it has been assigned. Figure 10 indicates that the performance of Openh4P and h4PI are very similar on the Origin 2000, right up to 128 processors. OpenMP is again slightly slower than the MPI implementation, but t,he timings differ by no more than 10% in all cases. The slightly slower OpenMP results may be due to the higher number of implicit synchrorli~atioils in this implementation.
2. Figure 13 depicts the relative timings for the same case rurlning in the hybrid MPI-OpenhlP mode on an Origin 2000, using different, combinations of h4PI and Openh/IP, up to 128 processors. for the threadto-thread communication model, while Figure> 14 dcpicts the results obtained using t,lie overlapping MPIOpenhlP communication model. The data point at, 64 MPI processes and 2 Openh4P threads in Figure 13 is not representative. since in this cast the opcrating system would occasionnlly place two threads on one processor. With this exception, both figures indicate that for small OpcrlMP th;td c:ount,s. R minor slowdown is observrd for the hybrid model over the pure MPI model. with mounting efficiency losses as the nurrib~r of threads is increased, although these are substantially smaller for the overlapping h4PI-Openh4P cominlillicatioii model.
3. Mp,6 8()-+- MPI 32 Speedups for two-level hybrid MPI-OpenMP Implementations of Single Grid Solver on SGI Origin 2000 using Thread-to-Thread Communication Strategy
4. 120 140 Fig. 14 Observed Speedups for two-level hybrid MPI-OpenMP Implementations of Single Grid Solver on SGI Origin 2000 using Overlapping Two-Level Communication Strategy