 |  |  | Future Work |
To provide temporal data, the system currently uses the context-switching among Java threads as the delimiter for counting intervals. Because the stop-the-world garbage collector prevents thread switching while it is executing, the system views a garbage collection, which can be longer than 10ms, as a single trace record. Because it is desirable to provide finer granularity for VM operations that disable context-switching, we plan to explore the addition of a separate trigger for capturing HPM information that will be in effect even when thread switching is disabled.
The adaptive optimization system of Jikes RVM's uses a cost/benefit model to determine which methods of an application should be optimized and at what optimization level [6]. We can leverage this system to attempt to selectively capture critical computations of the code in an automatic manner. Following in the direction of Arnold et al. [7], we could capture detailed profile information for only those methods that execute often. Specifically, we can instruct the optimizing compiler to automatically insert VM calls to partition the most time consuming computations and capture HPM information for these computations.
In this paper we explore using interactive visualization to drive performance analysis, which enables one to see large- and small-scale patterns, and to look at the data from any desirable perspective. A complimentary approach is to use statistical analysis to reduce the dimensional space. This is a worthwhile approach that we plan to explore.
| | | Future Work |