Title:
A Framework for Data Prefetching using Off-line Training of Markovian Predictors
A Framework for Data Prefetching using Off-line Training of Markovian Predictors
Authors
Kim, Jinwoo
Wong, Weng Fai
Palem, Krishna V.
Wong, Weng Fai
Palem, Krishna V.
Authors
Advisors
Advisors
Associated Organizations
Organizational Unit
Series
Collections
Supplementary to
Permanent Link
Abstract
An important technique for alleviating the memory bottleneck is data
prefetching. Data prefetching solutions ranging from pure software approach
by inserting prefetch instructions through program analysis to purely
hardware mechanisms have been proposed. The degrees of success of those
techniques are dependent on the nature of the applications. The need for
innovative approach is rapidly growing with the introduction of applications
such as object-oriented applications that show dynamically changing memory
access behavior. In this paper, we propose a novel framework for the use of
data prefetchers that are trained off-line. In particular, we propose two
techniques for building small prediction tables off-line and the hardware
support needed to deploy them at runtime. Our first technique is an
adaptation of the Hidden Markov Model that has been used successfully in
many diverse areas including molecular biology, speech, fingerprint and a
wide range of recognition problems to find hidden patterns. Our second
proposed technique is called the Window Markov Predictor, which seeks to
identify relationships between miss addresses within a fixed window. Sample
traces of applications are fed into these sophisticated off-line learning
schemes to find hidden memory access patterns and prediction models are
constructed. Once built, the predictor models are loaded into a data
prefetching unit in the CPU at the appropriate point during the runtime to
drive the prefetching. We will propose a general architecture for such a
process and report on the results of the experiments we performed, comparing
them against other hardware prefetching schemes. On average by using table
size of about 8KB size, we were able to achieve prediction accuracy of about
68% through our own proposed method and performance was boosted about 37% on
average on the benchmarks we tested. Furthermore, we believe our proposed
framework is amenable to other predictors and can be done as a phase of the
profiling-optimizing-compiler.
Sponsor
Date Issued
2002
Extent
736477 bytes
Resource Type
Text
Resource Subtype
Technical Report