Cache Line Boundary Allocation for Garbage Collected Systems

WAGLE, PRASAD AJIT. Cache Line Boundary Allocation for Garbage Collected Systems. (Under the direction of Dr. Edward F Gehringer.) Garbage-collected systems became increasingly popular with the release of the Java programming language. Cache performance of garbage-collected systems has been a heavily researched area. Past work has shown that cache-line utilization has been poor in garbagecollected systems. This work aims to reduce the cache miss rate by aligning objects to cacheline boundaries during object allocation. Object alignment on a cache line makes that object use minimum number of cache lines. This improves the overall utilization of a cache line between the time a memory block is brought into the cache and evicted from it. Improved cache utilization translates into improvement in total program execution time. Both cache performance (cache miss rate) and total execution time is studied on the Java Grande, DaCapo suite, Spec JBB2000, Spec JVM98 and soot benchmarks. The boundary allocation strategy is shown to improve cache performance/execution time for most memory-intensive benchmarks. The strategy performs well when implemented in the old generation of generational collectors. Boundary allocation in the young generation shows only marginal performance improvement. Most objects in the young generation “die” very soon. Hence the advantages of alignment do not outweigh the extra work of allocation. Performance of a benchmark will improve if fragmentation due to cache line alignment is diminished. It is shown that fragmentation depends on the object size distribution. Benchmarks that show performance improvement have a denser object-size distribution for object size equal to less than half the cache-line size and vice versa. It is also shown that fragmentation would decrease (hence performance would improve) with bigger cache lines. Cache Line Boundary Allocation for Garbage Collected Systems by Prasad Ajit Wagle A thesis submitted to the Graduate Faculty of North Carolina State University in partial fulfillment of the requirements for the Degree of Master of Science Computer Engineering Raleigh, North Carolina