Quantifying and improving the performance of garbage collection

Matthew Hertz, University of Massachusetts Amherst

Abstract

Researchers have devoted decades towards improving garbage collection performance. A key question is: what opportunity remains for performance improvement? Because the perceived performance disadvantage of garbage collection is often cited as a barrier to its adoption, we propose using explicit memory management as the touchstone for evaluating garbage collection performance. Unfortunately, programmers using garbage-collected languages do not specify when to deallocate memory, precluding a direct comparison with explicit memory management. We observe that we can simulate the effect of explicit memory management by profiling applications and generating object reachability information and lifetimes to serve as zero-cost oracles to guide explicit memory reclamation. We call this approach oracular memory management. However, existing methods of generating exact object reachability are prohibitively expensive. We first present the object reachability time algorithm, which we call Merlin. We show how Merlin runs in time proportional to the running time of the application and can be performed either on-line or off-line. We then present empirical results showing that Merlin requires orders of magnitude less time than brute force trace generation. We next present simulation results from experiments with the oracular memory manager. These results show that when memory is plentiful, garbage collection performance is competitive with explicit memory management, occasionally increasing program throughput by 10%. This performance comes at a price. To match the average performance of explicit memory management, the best performing garbage collector requires an average heap size that is two-and-a-half or five times larger, depending on how aggressively the explicit memory manager frees objects. This performance gap is especially large when the heap exceeds physical memory. We show the garbage-collector causes poor paging performance. During a full heap collection, the collector scans all reachable objects. Which a reachable object resides on a non-resident page, the virtual memory manager reloads that page and evicts another. If the evicted page also contains reachable objects, ultimately it too will be accessed by the collector and reloaded into physical memory. This paging is unavoidable without cooperation between the garbage collector and the virtual memory manager. We then present a cooperative garbage collection algorithm that we call bookmarking collection (BC). BC works with the virtual memory manager to decide which pages to evict. By providing available empty pages and, when paging is inevitable, processing pages prior to their eviction, BC avoids triggering page faults during garbage collection. When paging, BC reduces execution time over the next best collector by up to a factor of 3 and reduces maximum pause times by over an order of magnitude versus GenMS. This thesis begins by asking how well garbage collection performs currently and which areas remain for it to improve. We develop oracular memory management to help answer this question by enabling us to quantify garbage collection's performance. Our results show that while garbage collection can perform competitively, this good performance assumes the heap fits in available memory. Our attention thus focused onto this brittleness, we design and implement the bookmarking collector. While performing competitively when not paging, our new collector's paging performance helps make garbage collection's performance must more robust.

Subject Area

Computer science

Recommended Citation

Hertz, Matthew, "Quantifying and improving the performance of garbage collection" (2006). Doctoral Dissertations Available from Proquest. AAI3242341.
https://scholarworks.umass.edu/dissertations/AAI3242341

COinS