In Figures 9 and 10, we see that modeling error is high in
only a few cases (for example, for 64 aggregator core count
in Figure 9(a)). In addition, we observe that the overall
average percentage error was particularly high for Hopper
( 32% in Figure 9) but much less for Intrepid ( 20% in
Figure 9). These results lead us to conclude that Intrepid
is more well-behaved and simpler to model than is Hopper.
Similarly, in characterization study plots Hopper exhibited
higher variability in error than did Intrepid. One reason for
the dierence in behavior can be attributed to the fact that
the network phase, which is a component of the combined
throughput, is more well behaved on Intrepid and hence is
easier to model and predict. This is primarily to do with how
jobs are allocated by default on Blue Gene vs Cray machines.
On Blue Gene the communication trac is isolated, whereas
on Cray it can interfere with the trac of other jobs.