Characterizing and optimizing hotspot parallel garbage collection on multicore systems

The proliferation of applications, frameworks, and services built on Java have led to an ecosystem critically dependent on the underlying runtime system, the Java virtual machine (JVM). However, many applications running on the JVM, e.g., big data analytics, suffer from long garbage collection (GC) time. The long pause time due to GC not only degrades application throughput and causes long latency, but also hurts overall system efficiency and scalability. In this paper, we present an in-depth performance analysis of GC in the widely-adopted HotSpot JVM. Our analysis uncovers a previously unknown performance issue - the design of dynamic GC task assignment, the unfairness of mutex lock acquisition in HotSpot, and the imperfect operating system (OS) load balancing together cause loss of concurrency in Parallel Scavenge, a state-of-the-art and the default garbage collector in HotSpot. To this end, we propose a number of solutions to these issues, including enforcing GC thread affinity to aid multicore load balancing and designing a more efficient work stealing algorithm. Performance evaluation demonstrates that these proposed approaches lead to the improvement of the overall completion time, GC time and application tail latency by as much as 49.6%, 87.1%, 43%, respectively.