Portable Profiling of Memory Allocation in Java

This paper presents a portable framework to profile memory allocation in standard Java Virtual Machines. We extended our exact profiler JP, which generates a method call tree (MCT) for each thread in the system, in order to store information concerning object allocation in the MCT. Our primary design goals were to use platform-independent metrics for memory allocation and to minimize the extra overhead caused by memory profiling. For each method invocation context, the profiler preserves the number of allocated objects of each type. We exploit the fact that an object allocation is usually followed by a constructor invocation, in order to compute the number of object allocations from the MCT. A static analysis of constructor code allows to distinguish between the invocation of an alternate or superclass constructor and object allocation within the constructor. Arrays are treated specially, as we instrument array allocation instructions in order to preserve accumulated information on the type, number, and size of allocated arrays for each method invocation context. A performance evaluation shows that the extra overhead due to memory profiling is small.