ParaShares: Finding the Important Basic Blocks in Multithreaded Programs

Abstract

Abstract. Understanding and optimizing multithreaded execution is a significant challenge. Numerous research and industrial tools debug parallel performance by combing through program source or thread traces for pathologies including communication overheads, data dependencies, and load imbalances. This work takes a new approach: it ignores any underlying pathologies, and focuses instead on pinpointing the exact locations in source code that consume the largest share of execution. Our new metric, ParaShares, scores and ranks all basic blocks in a program based on their share of parallel execution. For the eight benchmarks examined in this paper, ParaShare rankings point to just a few important blocks per application. The paper demonstrates two uses of this information, exploring how the important blocks vary across thread counts and input sizes, and making modest source code changes (fewer than 10 lines of code) that result in 14-92% savings in parallel program runtime.

C oz

“…COZ identifies many lines in the CNDF and BlkSchlsEqEuroNoDiv functions that would have a small impact if optimized. This same code was identified as a bottleneck by ParaShares [26]; this is the only optimization we describe here that was previously reported. This block of code performs the main numerical work of the program, and uses many temporary variables to break apart the complex computation.…”

“…Techniques based on binary instrumentation can identify cache and heap performance issues, contended locks, and other program hotspots [5,31,36]. ParaShares and Harmony identify basic blocks that run during periods with little or no parallelism [25,26]. Code identified by these tools is a good candidate for parallelization or classic serial optimizations.…”

A fast causal profiler for task parallel programs

“…There is a large body of work to identify parallelism bottlenecks. These include techniques to address load imbalances [12,24,34,40], scalability bottlenecks [31,38,41], visualizing bottlenecks [13][14][15]25], synchronization bottlenecks [7,11,45], and data locality bottlenecks [1,28,30]. Data locality and synchronization bottlenecks increase serial work.…”

Parallel Execution Profiles

“…This approach essentially extends the gprof [5] style execution time profilers to handle multithreaded programs. [8] defines a metric called ParaShares as the normalized processor time with respect to weighted basic blocks. The key reason of using the above metrics for performance debugging is that the code which runs during periods of high parallelism has a smaller impact on overall execution time when optimized, than the code that runs during points of low parallelism.…”