Mimir: A Geometric Approach to Multi-dimensional Program Profiling Architectures

University Of California-Santa Barbara, Santa Barbara CA

Investigators

Abstract

CCF-0702798 Mimir: A Geometric Approach to Multi-dimensional Program Profiling Architectures Timothy P. Sherwood While mixed static-dynamic program analysis can be done completely in software through binary instrumentation, the amount of analysis that can be done at test-time is bounded by the performance impact that can be tolerated. The end goal of the Mimir project is to enable a new breed of hardware/software analysis tools, for researchers and system builders that can sift through on-line profile data at unprecedented speeds, yielding a highly accurate and timely image of computer system execution. The cross-layer approach to be investigated combines the raw computational ability of custom architectures with the formal guarantees provided by carefully crafted stream algorithms. At a high level, the proposed algorithmic approach to profiling is grounded in geometry, implicitly motivated by the belief that many profiling patterns, trends, or anomalies have natural geometric representations that become discernible under a geometric lens. At a low level, novel programmable hardware methods will provide a scalable and high performance substrate onto which these stream algorithms can be mapped. The combination of these two methods will allow online monitors to make streaming queries over live data at unprecedented speeds with the goal of enabling a new class of previously intractable dynamic analysis methods.