MUSA: A Multi-Level Simulation Approach for Next-Generation HPC Machines
SessionPerformance Tools
Session ChairLauren L. Smith
Event Type
Paper
Advanced
Intermediate
Performance
System Software
TimeWednesday, November 16th1:30pm - 2pm
Location355-E
DescriptionThe complexity of HPC infrastructures is growing due to a larger number of components and increasing heterogeneity. Interactions between software and hardware are not transparent to programmers and system architects. Therefore, predicting the behavior of applications on future systems is a challenging task.
In this paper, we present MUSA, an end-to-end methodology for multi-level simulation. Combining different levels of abstraction, MUSA models the communication network, microarchitectural details, and system software interactions, trading off simulation cost and accuracy. We compare detailed MUSA simulations with native executions of up to 2,048 cores and find errors to be less than 10% in the common case. Furthermore, we use MUSA to simulate up to 16,384 cores and identify scalability bottlenecks due to different factors - e.g. memory contention and load imbalance. We also compare different system configurations, showing how MUSA can help to assess the usefulness of future technologies in next-generation HPC machines.
In this paper, we present MUSA, an end-to-end methodology for multi-level simulation. Combining different levels of abstraction, MUSA models the communication network, microarchitectural details, and system software interactions, trading off simulation cost and accuracy. We compare detailed MUSA simulations with native executions of up to 2,048 cores and find errors to be less than 10% in the common case. Furthermore, we use MUSA to simulate up to 16,384 cores and identify scalability bottlenecks due to different factors - e.g. memory contention and load imbalance. We also compare different system configurations, showing how MUSA can help to assess the usefulness of future technologies in next-generation HPC machines.
Download PDF
Paper provided by the IEEE Computer SocietyPaper also available from the ACM Digital Library
Authors
Thomas Grass (presenting)
