Arvind Krishnamurthy, Jeanna Neefe, and Randy Wang
{arvindk,neefe,rywang}@cs.berkeley.edu
A parallel program, much like an elephant, is the organic union of
many complex parts. The
Splash-2 paper,
much like the blind men in the classic
Indian parable,
attempts to picture a complex object by studying its bits and pieces
in isolation. We show in this mini-project
(postscript),
that much like their
ancient blind predecessors, blanket statements made from a few poorly
chosen metrics tend to miss the whole picture and does little to help
the poor architect decide what machine to build to satisfy the needs
of these applications.
Introduction
Concurrency and Load Balance
Inverse Amdahl's Law
Effect of Single Cycle Memory Access Assumption
Synchronization Costs
Communication Behavior
Communication Granularity
Overlapping Communication and Computation
Traffic Burstiness
Summary
Analytical Model for Performance