White Papers

8
Peak Memory Consumption
The improvement in peak memory consumption is unusual given that an application typically uses a fixed amount of memory according
to input data size and/or the number of threads allocated. However, SOAPdenovo2 consumes 39.5% less memory on PowerEdge
R940 than PowerEdge R930 for 3.2 billion read data. Figure 6 shows the peak memory consumption over various data sizes with 62
cores on both systems. The reduction is relatively consistent over different data sizes as well as different number of cores used as
shown in Figure 7. The plot shows the peak memory consumptions with 31/62 cores on R930 and 31/62/110 cores on R940. The linear
trend lines are well fit onto the data points from the both systems.
One possible explanation for the peak memory reduction is larger memory bandwidth on R940. Although 2400MHz DIMMs are
populated in R930, the memory clock speed must be downgraded to 1333MHz due to the 3 DPC configuration which is theoretically
50% of the memory clock speed on R940. Actual memory bandwidth tests show more than 50% lower results for read/write on larger
data sets with loop unrolling codes as shown in Figure 8 and Figure 9. The bandwidth tests were performed with parallel memory
bandwidth (pmbw) benchmark/measurement (12).
Figure 6 Peak memory consumption with 62 cores. Peak memory
consumptions grow over the input data size.
Figure 7 Peak memory consumptions over various input data sizes and the number of
cores: the data points for R930 were generated by using 31/62 cores with 80/160/1600/3200
million reads data. For R940, the results with 31/62/110 cores for the same inputs were
plotted.