Towards Open-Source and Automatic Performance Characterization Hardware

By Matthew E. Weingarten, Columbia University in the City of New York. Tanvir Ahmed Khan, University of Michigan.

Abstract: Performance characterization is the key to unlocking efficient utilization of the underlying processing system. Rapid developments in specialized computing and hardware/software co-design make performance characterization more challenging—as the underlying hardware changes, so must the performance monitoring hardware and the accompanying performance models. CPU vendors have successfully popularized the top-down micro-architectural analysis (TMA) methodology that is effective in identifying true bottlenecks in a processor while abstracting away the hardware implementation. Unfortunately, researchers and practitioners are often limited to open-source RISC-V processors that lack hardware support for TMA or any other systematic performance characterization methodology. Even the simple scalar in-order RocketCore has not implemented performance hardware capable of providing enough information to support TMA, let alone more complex Out-of-Order (OoO) and super scalar SonicBOOM core. Furthermore, the challenges of performance characterization are compounded by the ever-increasing heterogeneity and specialization of hardware, and a wholistic performance characterization methodology for an entire System-on-Chip (SoC) remains open-ended. Overall, the lack of hardware-supported performance characterization hamstrings the ability to evaluate new hardware designs, for performance tooling to adapt to modern hardware, or even programmers efficiently exploit the target hardware.