Does the Halting Necessary for Hardware Trace Collection Inordinately Perturb the Results?

Processor address traces are invaluable for characterizing workloads and testing proposed memory hierarchies. Long traces are needed to exercise modern cache designs and produce meaningful results, but are difficult to collect with hardware monitors because microprocessors access memory too frequently for disks or other large storage to keep up. The small, fast buffers of the monitors fill quickly; in order to obtain long contiguous traces, the processor must be stopped while the buffer is emptied. This halting may perturb the traces collected, but this cannot be measured directly, since long uninterrupted traces cannot be collected. We make the case that hardware performance counters, which collect runtime statistics without influencing execution, can be used to measure halting effects. We use the performance counters of the Pentium 4 processor to collect statistics while halting the processor as if traces were being collected. We then compare these results to the statistics obtained from unhalted runs. We present our results in terms of which counters are affected, why, and what this means for trace-collection systems.

Name thesis:degree_name

MS

Grantor dc:publisher

Brigham Young University - Provo

Author dc:creator

• Watson, Myles G.

• performance counters
• processor hardware
• tracing address
• trace collection
• Computer Sciences

Language dc:language

English

Repository record dc:identifier

https://scholarsarchive.byu.edu/etd/201

OAI identifier oai:identifier

oai:scholarsarchive.byu.edu:etd-1200