Accurate and efficient memory monitoring in a Java virtual machine

Modern applications are often written in an object-oriented style. They allocate large amounts of objects and rely on automatic garbage collection to reclaim them again. Memory-related problems such as out-of-memory exceptions, long garbage collection pauses or performance degradation due to excessive garbage collection activity are hard to find without proper tools. Common memory monitoring tools record the memory state of an application either by sampling (i.e., by periodically taking a snapshot of the memory state) or by instrumentation (i.e., by adding code snippets that record memory events such as object allocations or garbage collector activity). Sampling is often too coarse to track down intricate memory problems, while instrumentation often generates too much overhead. This thesis presents the design and implementation of AntTracks, a novel Java virtual machine that is able to record memory events at object-level with negligible distortion of the applications behavior. The design includes a novel format for memory events designed to be compact and GC-independent, enabling a generic post-processing tool to analyze the memory behavior without any internal knowledge about the GC algorithm that generated the trace. Also, we describe how to generate these events efficiently in just-in-time compiled code and how to increase the trace accuracy with VM-internal information. The thesis also addresses the fact that capturing events at object-level leads to traces that are too big be stored in their full length. It proposes and evaluates several techniques for keeping traces small and manageable. We also present novel techniques for parsing and analyzing such traces. To the best of our knowledge, our approach is the first that captures information at object-level and still only requires as much heap memory as the monitored application. The AntTracks tool is able to show an overview of the memory behavior throughout the applications lifetime. It can reproduce the heap for any point in time, enabling detailed analyses about the heap state. Furthermore, the tool can analyze the changes to the heap over time, enabling analyses about trends in the heap. We provide both a qualitative evaluation, showing that the AntTracks virtual machine only imposes very little overhead compared to other state-of-the-art approaches, as well as a functional evaluation, showing how the recorded information can be used to track down memory-related performance problems.submitted by Dipl.-Ing. Philipp LengauerAbweichender Titel laut Übersetzung der Verfasserin/des VerfassersZusammenfassung in deutscher SpracheUniversität Linz, Dissertation, 2017OeBB(VLID)187309