Memoised Garbage Collection for Software Model Checking

摘要

Virtual machine based software model checkers like JPF and MoonWalker spend up to half of their verification time on garbage collection. This is no surprise as after nearly each transition the heap has to be cleaned from garbage. To improve this, this paper presents the Memoised Garbage Collection (MGC) algorithm, which exploits the (typical) locality of transitions to incrementally perform garbage collection. MGC tracks the depths of objects efficiently and only purges objects whose depths have become infinite, hence unreachable. MGC was experimentally evaluated via an implementation in our model checker MoonWalker and benchmarks using the parallel Java Grande Forum benchmark suite. By using MGC, a performance increase up to 78% was measured over the traditional Mark & Sweep implementation.