Headroom-based pretenuring: dynamically pretenuring objects that live "long enough"

Abstract

Many modern garbage collectors are generational, operating under the assumption that "most objects die young." Such collectors allocate all objects in a frequently collected nursery and tenure only surviving objects to a less-frequently collected, older generation. But these survivors induce copying costs upon tenuring. To avoid these costs, pretenuring schemes construct classifiers to predict whether an object will be long-lived or short-lived; accordingly, it is tenured or not immediately upon allocation. Up to now, however, these predictions did not account for one important fact: the proximity of the next collection. In contrast, headroom-based pretenuring does take this into account; thus, it can dynamically pretenure objects whenever they live "long enough." We devised two ways to estimate an object's lifetime from garbage collection traces. This led to two headroom-based pre-tenuring schemes, which we implemented on top of Jikes RVM and MMTk. Our experiments show that the dynamic, headroom-based pretenuring schemes outperform static schemes in terms of collector performance, albeit at the cost of increased mutator overhead.