Characterization of parallel application runtime sensitivity on multi-core high performance computing systems
Abstract
A commonly seen behavior of parallel applications is that their runtime is influenced by network communication load. The way a parallel application is run in a network and the presence of other applications and processes in the network can contribute to a wide range of variations in the runtime. Therefore, in order to achieve consistent and optimal runtimes, it is important to understand and characterize the runtime sensitivity of parallel applications with respect to execution under the presence of network communication load. In this research, runtime sensitivities for various parallel applications were studied by applying additional network communication load. In particular, the focus was on the runtime sensitivity of parallel applications on a multi-core multi-processor (MCMP) system where less network switching and routing are involved compared to single-core single-processor machines. The objective of this work was to determine if a previously developed sensitivity model for single-core single-processor machines still holds good for multi-core machines. For this purpose, previously developed tools (PACE and PARSE) were used to perturb the communication sub-system while executing several parallel application benchmarks such as the NAS benchmarks and PSTSWM. Runtime variations of these parallel applications were studied, under a specific network communication load, for different test cases by changing computing core allocation. A 10-node 80-core cluster was used as the test bed for this research purpose. Several test cases were explored using a variety of core allocations (process locations) for the application under test (AUT) to simulate job scheduler fragmentation. To ensure statistical significance, several iterations (trial runs) were executed in each test case. Results indicate that the idea of application sensitivity to communication sub-system performance degradation holds for MCMP architectures.
Degree
M.S.
Advisors
Evans, Purdue University.
Subject Area
Computer Engineering|Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.