Cold Object Identification and Segregation using Page Protection and Profiling

Abstract

Managed runtimes like the Java Virtual Machine (JVM) provide automatic memory management. This frees the programmer from explicitly freeing allocated memory and avoids memory leaks. Garbage collectors (GC) periodically walk the heap and free unused memory. Some of the surviving objects are infrequently accessed, but the JVM has to account for these objects during each GC cycle-clearly an unnecessary overhead. Such objects can be categorized as cold and moved to a dedicated memory. Typically hot regions are GCed more frequently, thus fewer objects have to be processed during each GC cycle and hence performance increases. Furthermore, cold objects can be stored in cheaper and larger capacity memory, like NVRam, leaving more RAM for the hot objects of the application. The identification of such objects is a particularly difficult task. This paper explores two methods of determining the temperature of objects. First, a coarse grain approach using the operating system's page protection is investigated. Second, an a priori profiling of an application is used to determine which classes will more likely produce objects that are hot or cold. Initial experiments, with a prototype implementation, show that this approach can lead to performance increases.