Roland Wunderlich Carnegie Mellon University

Super-resolution Performance Trace Reconstruction from CPU Performance Counters

CPU performance counters are invaluable tools to gain visibility to the internal operations of modern microprocessors during performance debugging and tuning by programmers and hardware designers. Unfortunately, attempting to extract measurements from hardware performance counters at too fine a granularity can perturb execution behavior and introduce context-sensitive observation effects; thus measurement intervals can be no smaller than milliseconds, i.e., tens of millions of instructions in today's CPUs.

I will present my current research on performance counter super-resolution that provides dynamic performance data at resolutions beyond performance counter limits. Super-resolution entails repeatedly measuring program execution events at low-resolution to reconstruct high-resolution and low-noise performance traces that can be used to pinpoint performance bottlenecks.

Bio

Roland is a PhD candidate in Electrical and Computer Engineering at Carnegie Mellon University. He is advised by Prof. James Hoe. Roland is writing his thesis on performance counter super-resolution. He has also worked on projects involving computer architecture simulation sampling and rapid FPGA prototyping.