Performance of modern computers is tied closely to the effective use of cache because of the continually increasing speed discrepancy between processors and main memory. We demonstrate that generational garbage collection employed by a system with cache and scratchpad memory can take advantage of the locality of small short-lived objects in Java and reduce memory traffic by as much as 20% when compared to a cache-only configuration. Converting half of the cache to scratchpad can be more effective at reducing memory traffic than doubling or even quadrupling the size of the cache for several of the applications in SPECjvm98.
Carl S. Lebsack, J. Morris Chang