Hinted Collection

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission. Acknowledgement I would like to thank Martin Maas, Joel Galenson, and Krste Asanovic for early constructive criticism of the ideas that appeared in this thesis. Many members of the Parallel Computing Laboratory (ParLab) have been subjected to half-baked versions of these ideas; thank you for your patience and questions. I would like to thank my advisor George Necula for his support. Abstract Hinted Collection Garbage collection is widely used and has largely been a boon for programmer productivity. However, traditional garbage collection is approaching both practical and theoretical performance limits. In practice, the maximum heap size and heap structure of large applications are influenced as much by garbage collector behavior as by resource availability. We present an alternate approach to garbage collection wherein the programmer provides untrusted deallocation hints. Usage of deallocation hints is similar to trusted manual deallocation, but the consequence of an inaccurate hint is lost performance not correctness. Our hinted collection algorithm uses these hints to identify a subset of unreachable objects with both better parallel asymptotic complexity and practical performance. We present two prototype implementations of a stop-the-world hinted collector: one entirely serial and one parallel. We evaluate our implementations by comparing against the Boehm-Demers-Weiser [12] conservative garbage collector for C/C++. We leverage existing free calls in mature C programs to stand in for deallocation hints. On some benchmarks, our serial collector implementation achieves 10-20% pause time reductions over a well-tuned baseline. On four cores, our parallel implementation achieves similar benefits. We include a discussion of the design trade-offs inherent in our approach, and lessons to be learned from our collectors. We close with a discussion of several design variants which we have not been able to explore in depth, but believe would be worthwhile to explore in future work.