JVM Configuration Management and Its Performance Impact for Big Data Applications

摘要

Big data applications are typically programmed using garbage collected languages, such as Java, in order to take advantage of garbage collected memory management, instead of explicit and manual management of application memory, e.g., dangling pointers, memory leaks, dead objects. However, application performance in Java like garbage collected languages is known to be highly correlated with the heap size and performance of language runtime such as Java Virtual Machine (JVM). Although different heap resizing techniques and garbage collection algorithms are proposed, most of existing solutions require modification to JVM, guest OS kernel, host OS kernel or hypervisor. In this paper, we evaluate and analyze the effects of tuning JVM heap structure and garbage collection parameters on application performance, without requiring any modification to JVM, guest OS, host OS and hypervisor. Our extensive measurement study shows a number of interesting observations: (i) Increasing heap size may not increase application performance for all cases and at all times, (ii) Heap space error may not necessarily indicate that heap is full, (iii) Heap space errors can be resolved by tuning heap structure parameters without enlarging heap, and (iv) JVM of small heap sizes may achieve the same application performance by tuning JVM heap structure and GC parameters without any modification to JVM, VM and OS kernel. We conjecture that these results can help software developers of big data applications to achieve high performance big data computing by better management and configuration of their JVM runtime.