Uppsala Architecture Research Team
StatCache/StatStack Statistical Cache Modeling
Statistical cache modeling is a collection of techniques for rapidly modeling the miss ratio of a cache or hierarchy of caches from low-overhead sparse data collected at runtime. StatCache and StatStack are novel sampling-based methods for performing data-locality analysis on realistic workloads. They are based on probabilistic models of the cache, rather than a functional cache simulator. The models use statistics from a single application run to accurately estimate miss ratios of fully-associative random and LRU caches of arbitrary sizes and generate working-set (miss ratio as a function of cache size) graphs. StatCache and StatCC have been evaluated using the SPEC benchmarks and shown to gives accurate results with a sampling rate as low as 10^(-4). This technology was commercialized as part of the ThreadSpotter(TM) tools from RogueWave.
Statistical Cache Modeling first samples architecturally independent reuse distances and then uses those to model behavior on arbitrary cache hierarchies.
- Efficient cache modeling with sparse data. In Processor and System-on-Chip Simulation, pp 193-209, Springer, New York, 2010. (DOI).
- StatStack: Efficient modeling of LRU caches. In Proc. International Symposium on Performance Analysis of Systems and Software: ISPASS 2010, pp 55-65, IEEE, Piscataway, NJ, 2010. (DOI).
- StatCache: A Probabilistic Approach to Efficient and Accurate Data Locality Analysis. In 2004 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS-2004),, 2004.
- StatCache: A Probabilistic Approach to Efficient and Accurate Data Locality Analysis. In Proceedings of the 2004 IEEE International Symposium on Performance Analysis of Systems and Software, 2004.