Hardware Reference
In-Depth Information
Chapter 26
Integrated Performance Monitoring
David Skinner
National Energy Research Scientific Computing Center, Lawrence Berkeley
National Laboratory
26.1
Design and Features ::::::::::::::::::::::::::::::::::::::::::::: 297
26.2
Success Stories ::::::::::::::::::::::::::::::::::::::::::::::::::: 301
26.2.1
Chombo's ftruncate :::::::::::::::::::::::::::::::::::::: 301
26.2.2
MADBENCH and File System Health :::::::::::::::::: 302
26.2.3
Buer Size ::::::::::::::::::::::::::::::::::::::::::::::: 303
26.2.4
HPC Workload Studies :::::::::::::::::::::::::::::::::: 304
26.3
Conclusion :::::::::::::::::::::::::::::::::::::::::::::::::::::::: 305
Bibliography :::::::::::::::::::::::::::::::::::::::::::::::::::::: 306
Application performance is usually tested as part of a benchmark or con-
trolled characterization study. Conducted by tools experts, the applications
are assessed in an \o-line" or non-production environment. In practice [6],
the performance delivered by an application over the duration of many execu-
tions can vary in uncontrolled and significant ways. A properly load-balanced
code and input deck can be easily, and often severely, bottle-necked by mi-
nor changes in input, concurrency, and execution environment. Many of these
\minor" changes are too easy to arrive at through natural adjustments of
scientific need or resource availability, which is the motivation behind the In-
tegrated Performance Monitoring (IPM) framework to measure and improve
in-vivo HPC application performance as it happens in a production computing
setting.
26.1 Design and Features
The IPM framework tests application performance in a production-like
environment to assesses application performance realistically and accurately.
IPM has been used in the context of individual work on codes [1], for HPC
system procurements [9], and metering workloads' sustained performance at
297
