Georg Hager's Blog

Random thoughts on High Performance Computing



Full-day tutorial at the Supercomputer Conference 2016 (SC 16), November 13-18, 2016, Salt Lake City, UT, USA:

Node-Level Performance Engineering

Slides: sc16_tutorial_nlpe_handout_update.pdf

Modified STREAM source code: stream.c

Compile with, e.g.:

icc -Ofast -xHost -qopenmp -fno-alias -nolib-inline -qopt-streaming-stores never|always -o stream.exe stream.c

Run with:

likwid-pin -c <pin_mask> ./stream.exe

LIKWID-instrumented STREAM source code: stream-mapi.c

Compile with, e.g.:

icc <options-from-above> -DLIKWID_PERFMON -I<path_to_likwid_inc> stream-mapi.c -o stream-mapi.exe -L<path_to_likwid_lib> -llikwid

Run with:

likwid-perfctr -C <pin_mask> -m -g <perf_group> ./stream-mapi.exe

Vector triad throughput benchmark: triad-throughput.tar.gz

Compile with:

icc -c timing.c
icc -c dummy.c
ifort -Ofast -xHost -qopenmp -fno-alias -fno-inline triad-tp.f90 dummy.o timing.o -o triad.exe

Run with:

echo <size> | likwid-pin <PIN_OPTIONS> ./triad.exe

Jacobi 3D stencil code: j3d_with_likwid.tar.gz

Build with the supplied Makefile (may need to adapt to your LIKWID setup).

Run with:

likwid-perfctr -C <pin_mask> -m -g <perf_group> ./J3D.exe <size>


Georg Hager1, Jan Eitzinger1 (not presenting), and Gerhard Wellein2

1 Erlangen Regional Computing Center
2 Department of Computer Science
University of Erlangen-Nuremberg



The advent of multi- and manycore chips has led to a further opening of the gap between peak and application performance for many scientific codes. This trend is accelerating as we move from petascale to exascale. Paradoxically, bad node-level performance helps to “efficiently” scale to massive parallelism, but at the price of increased overall time to solution. If the user cares about time to solution on any scale, optimal performance on the node level is often the key factor. We convey the architectural features of current processor chips, multiprocessor nodes, and accelerators, as far as they are relevant for the practitioner. Peculiarities like SIMD vectorization, shared vs. separate caches, bandwidth bottlenecks, and ccNUMA characteristics are introduced, and the influence of system topology and affinity on the performance of typical parallel programming constructs is demonstrated. Performance engineering and performance patterns are suggested as powerful tools that help the user understand the bottlenecks at hand and to assess the impact of possible code optimizations. A cornerstone of these concepts is the roofline model, which is described in detail, including useful case studies, limits of its applicability, and possible refinements.