ENZO

ENZO Performance Study Summary

On this page we will show the performance result for ENZO (svn repository version). We choose this version in part to see the effect of load balancing (not enabled in version 1.5) on scaling performance. The previous performance results for ENZO version 1 are at here.

Enzo Version 1.5

Following the release of Enzo 1.5 in November 08 we have done some follow up performance studies. Our initial findings are similar to what we found for version 1.0.1. For example, see this chart showing the scaling behavior of Enzo 1.5 on Kraken:

Scaling behavior was very similar on Ranger.

This poor scaling behavior could be anticipated by looking at the runtime breakdown (mean of 64 processors on Ranger):

With this much time spent in MPI communication increasing the number of processors above 64 is unlikely to result in much lower total execution time. Looking more closely at MPI_Recv and MPI_Barrier we see that on average 5.2ms is spent per call in MPI_Recv and 40.4ms for MPI_Barrier. This is much longer than can be explained by communication latency on Ranger's infiniband interconnect. Mostly likely then ENZO is experiencing a load imbalance causing some processors to wait for others to enter the Barrier or MPI_Send.

Next we looked at how enabling load balancing effect performance. This a runtime comparison between non-load balanced vs. load balanced simulation:

Time spent MPI_Barrier decrease but is mostly offset by the increase in time spent in MPI_Recv.

Callpath profiling gives us an idea where most of the costly MPI communications are taking place.

MPI Barriers take place in EvolveLevel(). And MPI_Recv takes place in grid::CommunicationSendRegions().

Snapshot profiles

Additionally, we used snapshot profiling to get a sense of how the ENZO's performance changed over the course of the entire execution. A snapshot was taken at each load balancing step such that each bar represents a single phase of ENZO between two load balancing phases. The first thing to notice is that these phases are regular and brief at the beginning of the simulation becoming progressively more varied with some becoming much longer.

(The time spent before that first load balancing has been removed--mostly initialization)

For MPI_Recv:

For MPI_Barrier: