hide
Free keywords:
Intel MPI, Asynchronous progress, Communication-computation overlap, Stencil code, Domain decomposition, Electronic structure, Octopus
Abstract:
When scaling HPC applications to large-scale systems, the
time spent in communication often becomes a bottleneck. A well-known
technique to tackle this problem is overlapping communication and
computation to hide communication time. In MPI codes, however, using
non-blocking functions is not enough -- the progress of the
communication needs to be triggered explicitly, either by the
application code or by special features of the MPI library. In this
talk, we explore overlapping communication and computation using the
asynchronous progress control feature of the Intel(r) MPI library by
applying it to stencil codes with a domain decomposition. With this
feature, the MPI library transparently handles the progress of
non-blocking MPI communication, removing the need for an explicit
control in the application. First, we introduce the asynchronous
progress control of the Intel(r) MPI library and how it can be used to
improve performance and scalability of a simple domain-decomposition
code. Second, we show how a real-world application can benefit from
this feature: the electronic structure code Octopus that uses finite
difference stencils to solve the time-dependent DFT equations.
Moreover, we try to generalize the conditions under which other
stencil codes with a domain decomposition can benefit as well. All
tests have been run on Cobra, the current flagship system of the Max
Planck Society with about 3400 nodes (having 2 Intel(r) Skylake 6148
Gold sockets each) and an Omnipath interconnect, hosted at the Max
Planck Computing and Data Facility (MPCDF).
