17.op/AJS.tockey
.ls 2
.na
.LP
Disk Caching

Barbara Tockey Zivkov

(Professor A. J. Smith)

Digital Equipment,
IBM,
MICRO, and
(NSF) CCR-91-17028

We are investigating the use of semiconductor memories as
high-speed caches for disk data in computer memory hierarchies.
As CPU speeds increase at a greater rate than disk I/O speeds,
the so-called "access gap" widens, and maximizing CPU utilization
depends more on the efficient management of disk data.

We have collected traces of disk I/O activity from computer
systems of varying workloads, from technical computing in a
research environment to production use of large databases
in a commercial environment. To our knowledge, no one
has amassed data from such a broad range of computing
systems for analysis in this manner.

We are analyzing this data, characterizing the I/O activity
to compare the behavior of these apparently very different
systems with respect to aspects such as locality of reference
and sequentiality. We are using trace-driven simulation to
study the effectiveness of common disk caching policies and
cache configurations, reporting results by such measures as
miss ratio and traffic ratio. We have simulated well-known disk caching
approaches such as least recently used and working set. We have
also simulated previously proposed policies which, for example,
dynamically determine the amount to pre-fetch, and have extended
these policies, for example, by determining pre-fetched amount on a per-file,
per-transaction, and per-file/transaction pair. We have simulated
the effects of limiting a transaction's allocation of cache buffers
and have investigated the effects of write policy on the traffic ratios.