Spelling suggestions: "subject:"diss date curves"" "subject:"diss date purves""
1 |
Generating Miss Rate Curves with Low Overhead Using Existing HardwareWalsh, Tom 17 February 2010 (has links)
Miss Rate Curves (MRCs) for main memory have been proposed as a representation of memory utilization for use in a range of optimizations in the area of memory man- agement. Various techniques exist for their creation; however, all real-world methods of MRC generation must make trade-offs between overhead and accuracy. Proposals for new hardware techniques exist, but have yet to be implemented in actual hardware. We pro- pose the use of the Intel PEBS (Precise Event-Based Sampling) performance monitoring capability for the task of MRC generation on existing commodity hardware.
We use PEBS to generate MRCs and compare them against MRCs generated through instrumentation, finding the PEBS MRCs to be good, but imperfect approximations, while keeping average PEBS overheads below 5%. We were unable to show that PEBS is better or worse than existing techniques, but believe we have succeeded in showing the promise of the use of general purpose performance monitoring hardware for this task and in motivating future research and development in this area.
|
2 |
Generating Miss Rate Curves with Low Overhead Using Existing HardwareWalsh, Tom 17 February 2010 (has links)
Miss Rate Curves (MRCs) for main memory have been proposed as a representation of memory utilization for use in a range of optimizations in the area of memory man- agement. Various techniques exist for their creation; however, all real-world methods of MRC generation must make trade-offs between overhead and accuracy. Proposals for new hardware techniques exist, but have yet to be implemented in actual hardware. We pro- pose the use of the Intel PEBS (Precise Event-Based Sampling) performance monitoring capability for the task of MRC generation on existing commodity hardware.
We use PEBS to generate MRCs and compare them against MRCs generated through instrumentation, finding the PEBS MRCs to be good, but imperfect approximations, while keeping average PEBS overheads below 5%. We were unable to show that PEBS is better or worse than existing techniques, but believe we have succeeded in showing the promise of the use of general purpose performance monitoring hardware for this task and in motivating future research and development in this area.
|
Page generated in 0.0533 seconds