Intel iWARP Quantum ESPRESSO Performance Study
iWARP vs. Infiniband Performance:
Quantum ESPRESSO
Quantum ESPRESSO (opEn Source
Package for Research in Electronic
Structure, Simulation, and Optimization)
is an integrated suite of computer codes
for electronic-structure calculations and
materials modeling at the nanoscale. It
is based on density-functional theory,
plane waves, and pseudopotentials (both
norm-conserving and ultrasoft). It is freely
available under the terms of the GNU
General Public License*.
The package builds onto newly
restructured electronic-structure codes
(PWscf, PHONON, CP90, FPMD, Wannier)
that have been developed and tested
by some of the original authors of novel
electronicstructure algorithms—from Car-
Parrinello molecular dynamics to density-
functional perturbation theory—and
applied in the last twenty years by some
of the leading materials modeling groups
worldwide. The Quantum ESPRESSO
distribution consists of a “historical” core
set of packages and a set of plug-ins that
performs more advanced tasks.
Quantum ESPRESSO is an initiative of the
DEMOCRITOS National Simulation Center,8
(Trieste) and SISSA9 (Trieste), in collabora-
tion with the CINECA National Supercom-
puting Center in Bologna,[1] the Ecole
Polytechnique Fédérale de Lausanne,[2]
the Université Pierre et Marie Curie,[3]
Princeton University,[4] Massachusetts
Institute of Technology,[5] and Oxford
University.[6]
For more information, see the Quantum
ESPRESSO website.[7]
Test Scenario
To broadly test the performance of iWARP
and InniBand with Quantum ESPRESSO,
the following workloads were developed,
results for each of which are presented
later in this section:
• Ausurf112 (DEISA benchmark): Self-
consistent cycle of the surface of Ag made
of 112 atoms, using a cut-off of 25 Ry and
four kpoints (2x2x1).
• Al256 (1 kpoints): Two steps of molecular
dynamics of liquid Al made of 256 atoms
using a cutoff energy of 50 Ry and the
gamma point.
• Al256 (8 kpoints): Identical to the previ-
ous description except using eight kpoints
(2x2x2).
Conclusion
By providing realistic application perfor-
mance instead of micro-benchmark test
results, this report illustrates the danger
of relying solely on synthetic benchmark-
ing when evaluating networking options.
HPC workloads behave much differ-
ently than, for example, a half round-trip
latency test: using multiple connections in
a switched environment with non-uniform
I/O patterns.
The real-world application results shown
in this report show the viability of iWARP
Ethernet as an alternative to discrete, pro-
prietary fabrics for HPC workloads. The
fundamental advantages of a converged
Ethernet network combined with easier
IP-based management and native routing
capability make iWARP a compelling solu-
tion for HPC use cases.
Figure 1. Quantum Espresso iWARP versus InfiniBand* Performance-Testing Results
(lower y-axis figures are better)
0.00
20.00
40.00
60.00
80.00
100.00
120.00
1x4 1x8 2x4 4x8 8x8
Time (mins)
Cores (nodes x cores per node)
Infinband*
iWARP
0.00
20.00
40.00
60.00
80.00
100.00
120.00
1x4 2x4 4x4 1x8 2x8 4x8 8x8
Time (mins)
Cores (nodes x cores per node)
Infinband*
iWARP
0.00
100.00
200.00
300.00
400.00
500.00
600.00
700.00
1x8 2x4 2x8 4x8 8x8
Time (mins)
Cores (nodes x cores per node)
Infinband*
iWARP
Ausurf112
AI256 (1 kpoints)
AI256 (8 kpoints)
2
Intel® Ethernet 10 Gigabit iWARP Quantum ESPRESSO Performance Study
2