Introduction to the X10 Implementation of NPB MG

performance metric FP PERCLOCK 4 LATENCY 375 (cycles) BANDWIDTH 5.3 (bytes/cycle) Computation Cost (Class S) 3,272,372(cycles) Ideal Speedup Number of places 1 2 4 8 Bulk-synchronous 1 1.74 3.17 5.46 Overlap comm & comp 1.19 2.22 4.08 7.05 Overlap comm 1.19 2.35 4.51 8.33 Replace for with foreach 5382.2 27.6 32.8 37.3 Figure 4: The ideal speedups are computed for four cases. For each case, different number of places are used to run the benchmark (Class S). The abstract performance metrics are defined in the top table. The computation and communication costs are inserted in the code manually. The runtime computes the critical path length. The base case is the bulk-synchronous implementation (SPMD style). Then parallelism is added gradually. First, the communication and computation are overlapped in each stencil operation. Secondly, communication is overlapped in the procedure of updating ghost values. Lastly, each unordered for loop in every stencil operation is replaced with a foreach loop. Abstract performance metric FP PERCLOCK 4 LATENCY 37.5 (cycles) BANDWIDTH 5.3 (bytes/cycle) Computation Cost (Class S) 3,272,372(cycles) Ideal speedup Number of places 1 2 4 8 Bulk-synchronous 1 1.85 3.58 6.83 Overlap comm & comp 1.19 2.38 4.67 9.07 Overlap comm 1.19 2.43 4.83 9.55 Replace for with foreach 5382.2 53.0 77.3 112.3performance metric FP PERCLOCK 4 LATENCY 37.5 (cycles) BANDWIDTH 5.3 (bytes/cycle) Computation Cost (Class S) 3,272,372(cycles) Ideal speedup Number of places 1 2 4 8 Bulk-synchronous 1 1.85 3.58 6.83 Overlap comm & comp 1.19 2.38 4.67 9.07 Overlap comm 1.19 2.43 4.83 9.55 Replace for with foreach 5382.2 53.0 77.3 112.3 Figure 5: The setting is the same as in Figure 4 except that the communication latency is reduced by a factor of 10. We can see that communication latency is the performance bottleneck of this kind of application In the last case shown in Figure 4, the ideal speedup decreases dramatically from using 1 place to 2 places. The cause is the high communication-computation ratio, that is, the cost to partition computation across computing nodes (places) is relatively high. In Figure 5, the communication latency is reduced by a factor of 10. We can see that communication latency is the performance bottleneck of this kind of applications.