Summary: Communication Behavior of a Distributed Operating System
Remzi H. Arpaci
Department of Electrical Engineering and Computer Science
Computer Science Division
University of California, Berkeley
We present measurements of the communication behavior of a prototype distributed operating system, Solaris MC.
We employ three server workloads to drive our study: a build of the Solaris source tree, a synthetic web server, and a
parallel database. Our measurements reveal a number of facts, which have implications on design of Solaris MC, the
prototype implementation of Solaris MC, and the design of a message layer. We find that message traffic is centered
around nodes that house disks, a potential bottleneck. A file system that striped data across the cluster would avoid such
a problem. Most messages are medium sized, in the range of 68 to 256 bytes, indicating that messagelayer support
for such messages is crucial. Further, messages are structured as a chain of two to three buffers, perhaps suggesting
the need for a gather interface to avoid additional buffer allocation and memory copies. Finally, requestresponse time
is quite high, due to overhead of the current message layer; this fact is perhaps most indicative of the prototype status
of the system. In any case, a lowoverhead message layer would substantially improve overall performance.
1 Introduction 5