CYCLER: IMPROVE HEAP MANAGEMENT FOR ALLOCATION-INTENSIVE APPLICATIONS WITH ON-THE-FLY OBJECT REUSE

摘要

Garbage collection (GC) is an indispensable technology inrnmanaged runtime systems (e.g. Java VM). The folk belief isrnthat it is hard for any other GC algorithm to beat the performancernof current sophisticated copying generational GCrnon short-lived objects. In this paper we describe Cycler thatrncan improve the performance and scalability of allocationintensivernapplications over copying generational GC. Wernobserve that the state-of-the-art tracing GCs reclaim thernspace of dead objects all-at-once in batch mode, so the deadrnobjects between two GC cycles are not collected. If there arerna large number of objects dying in the duration, the nonrecycledrnspace causes high cache miss rate and puts memoryrnbus under heavy pressure. The circumstance gets worse onrnmulti-core platforms due to the competition over memoryrnand bus subsystems. We propose a new GC design Cyclerrnthat enhances tracing GC with on-the-fly object reuse. Cyclerrncan effectively reduce GC pause time and improvernmemory access performance for allocation-intensive applications.rnThe novelty of Cycler is that it can be seamlesslyrnintegrated with existing tracing GCs without sacrificing thernlatter’s advantages.