Optimizing Trace Tool-overhead for Lock-Intensive Multi-threaded Parallel Applications

Abstract

Often a tool collecting traces for lock-intensive applications adds overheads of its own and distorts the lock-related measurements. We highlight why tool-overhead is particularly problematic for lock-intensive applications. Tool-overhead has received limited attention in existing research. The primary reason for high tool-overhead, as per our analysis is cache- coherence related overheads for tracing tool data structure. Using the insight, we develop Mutexis, an optimized user-level dynamic binary instrumentation (DBI) tracing PIN tool. To show tool effectiveness, we use lock-intensive applications from PARSEC and Splash3X benchmarks. We compare the proposed tool's overhead with tool-overhead of other researchers. The tool-overhead of mutexis is minimal, growing up to 2.1X for lock- intensive applications (4X to lOOX for others) and is negligible in most cases. This is so, even when our tool captures detail cycle- stamped traces of POSIX lock function compared to limited aggregate statistics collected by other researchers tools.