 |  |  | Accessing Hardware Counters |
Several library packages provide access to hardware performance counter information, including the HPM toolkit [15], PAPI [11], PCL [9], and OProfile [26]. These libraries provide facilities to instrument programs, record hardware counter data, and analyze the results. We extend the functionality of existing libraries to obtain hardware performance data in a virtual machine. Specifically, we extend Jikes RVM to collect thread-specific, temporally fine-grained hardware counter data in an SMP environment.
The Digital Continuous Profiling Infrastructure provides a powerful set of tools to analyze and collect hardware performance counter data on Alpha processors [5]. The system includes tools to collect accurate profile data with very low overhead and to analyze the profile data using many performance metrics. The system works on unmodified executables on multiprocessors and collects time-based hardware counter samples of program counter values. VTune [32] and SpeedShop [35] are similar tools from Intel and SGI, respectively. Our work differs in that we are interested in correlating the hardware counter data to high-level program constructs, such as Java threads, to distinguish the effects from the VM and user applications, in an SMP environment in a temporal manner.
Ammons et al. [4] correlate hardware performance counter information to frequently executed program paths. They use flow- and context-sensitive data-flow analysis techniques to collect hardware counter data along program paths instead of just individual statements or procedures. Although this provides fine-grained information, the overhead of recording hardware counter data along the paths increases runtime by an average of 70%. The overhead of collecting and storing our HPM trace files is less than 2% in our VM.
IBM's Performance Inspector [30] is a collection of profiling tools. It includes a patch to the Linux kernel providing a trace facility and hooks to record scheduling, interrupt and other kernel events. The package contains tools for sampling-based profiling (any performance counter can be used to trigger sampling), manual instrumentation-based profiling (measuring per-thread time/performance counts), and exception-based basic block profiling. Our work profiles several performance counters at once, does not require instrumentation by hand, and is tightly integrated with our VM, resulting in access to information about the runtime system (like compilation and garbage collection).
| | | Accessing Hardware Counters |