Heng Ma, Austin Clyde, Anda Trifan, Venkatram Vishwanath, Arvind Ramanathan,
Debsindhu Bhowmik, and Shantenu Jha.
Benchmarking machine learning workloads in structural bioinformatics
applications.
First International Workshop on Benchmarking Machine Learning
Workloads on Emerging Hardware: CHALLENGE20, 2020.
MA2020-AT
Benchmarking machine learning (ML) based methods have traditionally
been largely uncoupled from scientific simulations. However, there
has been considerable interest in using learning approaches in the
context of scientific simulation software to: (1) analyze large volumes of
simulation data (or archived databases);(2) drive adaptive simulations
to sample ‘rare’ events; (3) accelerate simulations by replacing
expensive computational kernels with efficient ML inference techniques,
and (4) drive optimal simulation strategies based on ML guided
approaches. Thus, the coupling of learning with simulation tools can
range widely: from ML approaches which are independent of the
application itself, to ML approaches which are used to drive large-scale
simulations. Using structural bioinformatics applications, we motivate
how ML approaches are coupled with physics-based simulations. To
optimize such coupled applications on emerging hardware and
software platforms, we need to consider additional and often unique
performance considerations. In this paper, we present an overview
of different learning approaches in structural bioinformatics applications, performance considerations for
such coupled applications, and
outline the development of performance metrics. We hope this will
enable the broader scientific community as well as hardware and software
vendors to evaluate the role of learning tools when coupled to scientific
simulation applications, and hope that this could serve as a framework
for other application domains.
