RASCAL -- Timestamp-Based Selective Cache Allocation
Abstract
The behavior of the memory hierarchy is key to high performance in
today's GHz microprocessors. The cache level closest to the processor is
limited
in size and associativity in order to match the short cycle time of the
CPU. While a substantial amount of cache research has been devoted to
replacement policies that minimize the conflict misses caused by limited
associativity, little effort has been put into finding selective cache
allocation policies that treats the limited space of the L1 cache with
care. Even though only the data objects that are reused soon again will
benefit
from allocation in the small L1 cache, all data objects are allocated in
most of
the L1 cache implementations.
It has been shown that optimal allocation algorithms with knowledge
about the future can drastically cut the miss ratio by only allocating
data
objects that are reused before the objects currently residing in the L1
cache. In this paper we show that the effect of optimal allocation can
be
further enhanced by the introduction of a small filtering cache, K1,
used as a
staging buffer in addition to the L1. We further evaluate three options
for
practical selective cache allocation algorithms in combination with a
small
staging cache. We show that the RASCAL,Runtime Adaptive Selective Cache
ALlocation
algorithm, an algorithm based on time stamps, is a practical and
efficient option.
The RASCAL algorithm detects a good L1 object by monitoring the duration
between
two consecutive allocations of a cache block as measured in the new time
unit
cache allocation ticks, CAT. CAT is shown to be a fairly accurate and
application-independent way of detecting good allocation candidates. A
reduced
RASCAL implementation requires 8 extra bits of time stamp information
for each
64 byte cache block in the L2. Resulting in less than 2% SRAM overhead.
Given a
2KB streaming buffer the RASCAL algorithm reduces the miss ratio with
more than
40 %, for 5 out of 14 applications and performs better or comparable to
a
conventional cache with optimal allocation in 4 out of fourteen
applications.
Place: Information technology, Uppsala University
Room: 1113
Time: 15.15-16.00 (+ discussions)
Room 1113 is in building 1, floor 1, room 13
(in the southern part of the building).
Help on how to find ASTEC Seminars.
There will be an extended period for discussions after the seminar.
Speakers are encouraged to give an short (5 min) introduction to the subject at the begining of the talk.
Listeners are excused if they have to leave after 16.00.
Everyone is welcome !

Updated 06-Jun-2001 17:40 by Roland Grönroos
e-mail: info -at- astec.uu.se
Location: https://www.astec.uu.se/Seminars/01/0612.shtml
|