17.op/JKO.bakermg
.m From mgbaker@allspice.Berkeley.EDU Wed Aug 25 16:11:04 1993
.ls 2
.na
.LP
Fast Crash Recovery in Distributed File Systems

Mary Baker

(Professor J. K. Ousterhout)

(ARPA/NASA) NAG-2-591 and (NSF) CCR-89-00029

This project focuses on fast crash recovery for improving
distributed system availability. The traditional way to improve availability
is to use various forms of redundancy to mask failures; unfortunately, this
approach can result in higher cost, increased complexity, and reduced
performance. Using fast crash recovery, I assume that critical resources will
fail, but I design them to recover so quickly that nobody is inconvenienced.
This method is simple and inexpensive and has little performance overhead. In
particular, I examine the recovery of distributed state such as the file caching
information maintained on file servers.

Using the Sprite distributed file system, I compare the recovery speed,
performance overhead, and complexity of three different fast crash recovery
techniques for file servers. In the first technique, client-driven recovery,
clients of the file server send their distributed state information to the
server after a crash. In the second technique, server-driven recovery, the
server initiates recovery with each of the clients. The last technique,
transparent recovery, employs non-volatile memory on servers so that clients
and servers need not communicate at all during recovery. This technique
provides the fastest recovery, allowing a Sprite file server with 40 clients
to recover from a crash in 20 seconds.