Quantifying the Effects of Network Performance on End Applications
Performance

by Richard Martin, Amin Vahdat, and David Culler

Abstract:

Parallel computing platforms are more and more frequently being
designed from commodity hardware.  This approach has a number of
advantages, including: rapid development time, fully functional OS
(allowing compatibility with all existing applications), and
cost-effective computing.  This study will enable network designers
and system engineers to assess the relative importance of various
network characteristics on end application performance.

However, the network performance of such systems often lags behind
what is available from many custom platforms (for example, the T3E).
Our work attempts to quantify the impact of network performance on a
wide variety of parallel applications.  The initial benchmark suite
includes: Sort, Cholesky, Ray Tracer, Murphi, Em3d, Barnes Hut, and
Connected Components.  The programs will be run on a cluster of 32
Ultrasparcs interconnected by a Myrinet network.

Our methodology involves running a benchmark suite while varying
network characteristics as prescribed by the LogP model.  LogP models
network characteristics with the following parameters: latency,
overhead, and gap.  We vary these parameters by introducing delays
into strategic locations in the Generic Active Messages library (both
in the user-level and in code running on the network interface).