[Colloquium] Marincic/MS Presentation/Apr 30, 2018

[Margaret Jaffey via Colloquium]

This is an announcement of Ivana Marincic's MS Presentation.

------------------------------------------------------------------------------
Date:  Monday, April 30, 2018

Time:  2:00 PM

Place:  Eckhart 117

M.S. Candidate:  Ivana Marincic

M.S. Paper Title: Application-Level Power Monitoring and Management in
High Performance Computing

Abstract:
The next generation of exascale supercomputing systems is planned to
perform a billion billion operations per second to push the frontiers
of science. Achieving such compute power raises many new challenges
including two addressed by this thesis: operating within strict power
limits, and the widening gap between compute and IO speed. Future
systems are expected to handle and maximize performance under strict
power limits. Developing software that meets power and energy
constraints is a difficult task that must account for dynamic
application behavior and system properties simultaneously. Scientific
application developers therefore must have a reliable way to measure
power and energy consumption and relate these measurements to
application- and system-specific events. Currently, scientists face
two challenges when measuring and controlling power: (1)
diversity---power and energy measurement interfaces differ between
vendors---and (2) distribution---power measurements of distributed MPI
applications should be unaffected by how a particular application’s
MPI processes are scheduled. The current state-of-the-art requires
scientists interested in power optimization to write tedious,
error-prone application- and system-specific code. To make power
measurement and management easier for scientists, we propose PoLiMEr,
a user-space library that supports fine-grained application-level
power monitoring and capping. We evaluate PoLiMEr by deploying it on
Argonne National Laboratory's Cray XC40 Theta system and using it to
measure and cap power, scaling the performance and power of several
applications on up to 1024 nodes. We find that PoLiMEr requires only a
few additional lines of code, but easily allows users to detect energy
anomalies, apply power caps, and evaluate Theta's unique architectural
features. The remainder of this thesis relates the second exascale
issue -- IO speeds falling behind compute speeds --- to the issue of
power constrained environments. As a result of increasing quantities
of data produced by simulations, scientists must turn to online
analysis of output data from large-scale simulations. Numerous
frameworks and scheduling techniques of simulation and analysis
workloads have been proposed to address this challenge, but none
consider the impact of power caps. Using PoLiMEr, we investigate this
impact, and propose and implement a power allocation strategy that
manages co-dependent executions of simulation and analysis. We
evaluate our method on the molecular dynamics simulation LAMMPS on
Theta and show we can achieve average improvements of ~30% in runtime
under strict power constraints.

Ivana's advisor is Prof. Henry Hoffmann

Login to the Computer Science Department website for details:
 https://www.cs.uchicago.edu/phd/ms_announcements#imarincic

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey            margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 156)               (773) 702-6011
The University of Chicago      http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=