Intelligent Management of Inter-Thread Synchronization Dependencies for Concurrent Programs.

Abstract

Power dissipation limits and design complexity have made the microprocessor industry less successful in improving the performance of monolithic processors, even though semiconductor technology continues to scale. Consequently, chip multiprocessors (CMPs) have become a standard for all ranges of computing from cellular phones to high-performance servers. As sufficient thread level parallelism (TLP) is necessary to exploit the computational power provided by CMPs, most performance-aware programmers need to parallelize their programs.

For shared memory multi-threaded programs, synchronization mechanisms such as mutexes, barriers, and condition variables, are used to enforce the threads to interact with each other in the way the programmers intended. However, employing synchronization operations in both correct and efficient way at the same time is extremely difficult, and there have been trade-offs between programmability and efficiency of using synchronizations.

This thesis proposes a collection of works that increase the programmability and efficiency of concurrent programs by intelligently managing the synchronization operations. First, we focus on mutex locks and unlocks. Many concurrency bug detection tools and automated bug fixers rely on the precise identification of critical sections guarded by lock/unlock operations. We suggest a practical lock/unlock pairing mechanism that combines static analysis with dynamic instrumentation to identify critical sections in POSIX multi-threaded C/C++ programs. Second, we present Dynamic Core Boosting (DCB) to accelerate critical paths in multi-thread programs. Inter-thread dependencies through synchronizations form critical paths. These critical paths are major performance bottlenecks for concurrent programs, and they are exacerbated by workload imbalances in performance asymmetric CMPs. DCB coordinates its compiler, runtime subsystem, and architecture to mitigates such performance bottlenecks. Finally, we propose exploiting synchronization operations for better energy efficiency through dynamic power management.