Benchmarks

NERSC Application Benchmarks

Workload Profiles IO Benchmarks Trace Data

Workload Profiles

The links below show workload characterization data of scientific applications at NERSC using the IPM profiling utility on Seaborg . Similar data are also available for Bassi.
The profiles provide a performance summary of the computation and communication in these applications, and were created by the Science-Driven System Architecture Team as part of their ongoing effort related to measurement and analysis of the NERSC scientific computing workload.
IPM currently monitors the following:

MPI: communication topology and statistics for each MPI call and buffer size.
HPM: PAPI (many) or PMAPI (AIX) performance events.
Memory: wallclock, user and system timings
Switch: Communication volume and packet loss

GTC: 3D Gyrokinetic Toroidal Code
GTC is a 3-dimensional code used to study microturbulence in magnetically confined toroidal fusion plasmas via the Particle-In-Cell (PIC) method ...more>>>

IPM benchmarks: Medium and Large datasets.

Paratec: Parallel Total Energy Code
The benchmark code PARATEC (PARAllel Total Energy Code) performs ab-initio quantum-mechanical total energy calculations using pseudopotentials and a plane wave basis set. ...more>>>

IPM benchmarks: Medium and Large datasets.

MADBench: Microwave Anisotropy Dataset Computational Analysis Package Benchmark
The benchmark code MADBench is a "stripped-down" version of MADCAP, a Microwave Anisotropy Dataset Computational Analysis Package ...more>>>

IPM benchmarks: Medium, Large and X-large datasets.

MILC: MIMD Lattice Computation
The benchmark code MILC represents part of a set of codes written by the MIMD Lattice Computation (MILC) collaboratoration used to study quantum chromodynamics (QCD), the theory of the strong interactions of subatomic physics ...more>>>

IPM benchmarks: Medium and Large datasets.

PMEMD: Particle Mesh Ewald Molecular Dynamics
The benchmark code PMEMD (Particle Mesh Ewald Molecular Dynamics) performs various molecular simulations including Molecular Dynamics (MD), NMR Refinement and minimizations ...more>>>

IPM benchmarks: Medium and Large datasets

IO Benchmarks with IPM*

The new version of IPM integrates the standard POSIX IO calls and can be used to trace I/O behavior of HPC applications.The following runs are made with this new feature on Jacquard (courtesy of Noel Keen)

MADBench:

256 tasks, POSIX one file per task [plots] [stats]
64 tasks, POSIX one file per task [plots] [stats]
16 tasks, POSIX shared file [plots] [stats]

Chombo:

256 tasks, 2 components [plots] [stats]
32 tasks, 2 components [plots] [stats]
32 tasks, 10 components [plots] [stats]

AMRScalingXfer: 128 tasks, small run [plots] [stats]

*Note: This is development software, and the runs/plots aren't well documented -- it's just a demonstration of the possibilities of doing IO profiling in IPM.

Trace Data

Here are files containing trace data for some of the applications. These traces are generated by invoking the "strace" utility on every task and piping the data for each task to a separate file. Process ID is used to create unique file names. All applications where run on Jacquard . The files are compressed tar files of the trace data

PMEMD 16 tasks, small dataset run

MADbench 64 tasks, medium dataset run

MILC 16 tasks, medium dataset run