Home Publications edited volumes Awards Research Teaching Miscellaneous Full CV [pdf] BLOG
Events
Past Events
|
Publications of Torsten Hoefler
Torsten Hoefler:
| | Characterizing the Influence of System Noise on Large-Scale Parallel Applications
(Presentation - presented in Aachen, Germany, Apr. 2011, Talk at RWTH Aachen University )
AbstractSystem noise is increasingly a concern as HPC systems continue
to grow in scale. Good operating systems can minimize noise, however,
Some sources of asynchronous slowdowns, such as recoverable hardware
error remain. Existing studies with artificial noise models provide
only limited insight into application behavior under the influence of
noise. This paper presents an in-depth analysis of the impact of
system noise on large-scale parallel application performance in
realistic settings. Our analytical model shows the particular
circumstances under which noise is propagated or absorbed. The model
shows that not only collective operations but also point-to-point
communications influence the application's sensitivity to noise. We
present a simulation toolchain that injects noise delays from traces
gathered on four common large-scale architectures into a LogGPS
simulation and allows new insights into the scaling of applications in
noisy environments. Our simulation framework enables large-scale
simulations up to 8 million processes with more than 1 million events
per second. We investigate collective operations in noisy settings
with up to 1 million processes and three applications (Sweep3D, AMG,
and POP) with up to 32.000 processes. We show that the scale at which
noise becomes a bottleneck is system-specific and depends on the
structure of the noise. Simulations with different network speeds
show that a 10x faster network does not improve application
scalability because noise becomes a bottleneck at scale. We quantify
this noise bottleneck and conclude that our tools can be utilized to
tune the noise signatures of a specific system for minimal noise
propagation. For example, our simulations verify the long-standing
conjecture that co-scheduling prevents significant application
slowdown.
Documents download slides: | | BibTeX | @misc{osnoise-talk-aachen, author={Torsten Hoefler}, title={{Characterizing the Influence of System Noise on Large-Scale Parallel Applications}}, year={2011}, month={Apr.}, location={Aachen, Germany}, note={Talk at RWTH Aachen University}, source={http://www.unixer.de/~htor/publications/}, } |
|
|