Title: "Performance Availability in Cluster I/O"

Speaker: Remzi H. Arpaci-Dusseau, U.C. Berkeley



Abstract:

In this talk, I will discuss how to achieve common-case peak performance
for I/O-intensive cluster applications. The motivation for this work is
experience with NOW-Sort, a high-performance external sort for clusters
of workstations. NOW-Sort attains peak performance only at night -- when
machines are otherwise idle, and all potential performance distractions
are manually removed from the system. However, when run in a less
sterile (and more realistic) environment, performance suffers noticeably.

The main reason for this lack of "performance availability" is the
presence of performance anomalies in clustered systems.  Due to the
complexity of both hardware and software, the behavior of machines
across a seemingly homogeneous pool of machines is often quite
varied. Software predicated on this homogeneity will exhibit erratic
performance, often an order of magnitude worse than expected. To remedy
this, systems must assume that such performance variations exist, and
contain provisions to operate well in spite of them. Towards this end,
I will describe a system called River. River employs two key ideas to
avoid performance faults: load balancing, via high-performance
distributed queues, and replication, with graduated declustering.  With
these two components, River facilitates the construction of applications
that perform gracefully even under severe performance anomalies,
allowing data to flow seamlessly around such faults. The result is
nearly ideal performance at all times -- whether day or night -- with
applications effectively utilizing their share of resources.


Biography

Remzi H. Arpaci-Dusseau is currently a graduate student at
U.C. Berkeley, under advisor David Patterson. He received a B.S. in
Computer Engineering, summa cum laude, from the University of Michigan
in 1993, and a Masters in Computer Science from U.C. Berkeley in
1996. He plans to complete his dissertation work in the fall of 1999.
His interests lay largely in the area of experimental distributed and
parallel systems, including operating systems, file systems, databases,
and computer architecture. His most recent work has been on River, a
software system designed to provide consistent, high-performance for
cluster applications with large I/O demands. He and his wife, Andrea
Arpaci-Dusseau, broke and still hold two world records in external 
sorting.   

For more information, see:
http://www.cs.berkeley.edu/~remzi