Optimizing Memory Management in Virtual Machine

Abstract

Memory management is one of key components in virtual machine and also affects overall performance of virtual machine itself. Modern programming languages for virtual machine use dynamic memory allocation and objects are allocated dynamically to heap at a higher rate, such as Java. These allocated objects are reclaimed later when objects are not used anymore to secure free room in the heap for future objects allocation. Many virtual machines adopt garbage collection technique to reclaim dead objects in the heap. The heap can be also expanded itself to allocate more objects instead. Therefore overall performance of memory management is determined by object allocation technique, garbage collection and heap management technique. In this paper, three optimizing techniques are proposed to improve overall performance of memory management in virtual machine. First, a lazy-worst-fit object allocator is suggested to allocate small objects with little overhead in virtual machine which has a garbage collector. Then a biased allocator is proposed to improve the performance of garbage collector itself by reducing extra overhead of garbage collector. Finally an ahead-of-time heap expansion technique is suggested to improve user responsiveness as well as overall performance of memory management by suppressing invocation of garbage collection. Proposed optimizations are evaluated in various devices including desktop, embedded and mobile, with different virtual machines including Java virtual machine for Java runtime and Dalvik virtual machine for Android platform. A lazy-worst-fit allocator outperform other allocators including first-fit and lazy-worst-fit allocator and shows good fragmentation as low as rst-t allocator which is known to have the lowest fragmentation. A biased allocator reduces 4.1% of pause time caused by garbage collections in average. Ahead-of-time heap expansion reduces both number of garbage collections and total pause time of garbage collections. Pause time of GC reduced up to 31% in default applications of Android platform.