(A) phase-adaptive garbage collector for java virtual machine

Abstract

Languages such as Java and C# rely on garbage collection (GC) to manage their memory. GC relieves developers from error-prone memory reclamation but application performance is hurt by the additional GC work. To reduce GC overhead, most garbage collectors uses a simple heuristic. The same heuristic is applied throughout the entire execution time, regardless of the kind of memory objects and the kind of applications. However, the memory behavior of applications is significantly different. Moreover, even a single application may show very different behaviors as it executes. Therefore, the traditional garbage collectors fail to support a wide range of applications and to match the varying behavior of an application. This thesis proposes a phase-adaptive garbage collector (PAGC) for Java virtual machine. The goal of the proposed collector is to improve GC efficiency by tailoring all its operations to the behavioral pattern (i.e., the phase) of applications. To this end, PAGC dynamically identifies application phases by monitoring application methods. Then, PAGC groups memory objects by their associated phase and determine the best GC timing for the grouped objects by considering the phase behavior. Furthermore, to collect the determined group efficiently, PAGC uses either the in-place GC or the phase-boundary GC. Applications usually have a ramp phase that objects are intensively allocated and a cliff phase that a lot of objects die en masse. The in-place GC and the phase-boundary GC improve the efficency of the garbage collection invoked at the ramp phase and at the cliff phase respectively. We implemented PAGC on SableVM research virtual machine and evaluated it using several benchmarks. The experimental results show that PAGC successfully recognizes application phases. Compared with the traditional generational garbage collector, PAGC improves the garbage collection time by as much as 41%.