Embracing heterogeneity with dynamic core boosting

Abstract

Uniformly distributing parallel workloads amongst threads is an effective strategy for programmers to increase application performance. However, in any parallel segment, execution time is determined by the longest running thread. Even for embarrassingly parallel programs in the form of SPMD (single program multiple data), the threads are not perfectly balanced due to control flow divergence, non-deterministic memory latencies, and synchronization operations. Such an imbalance can be significantly exacerbated by performance asymmetry among cores, which is likely to exist in future generations of chip multiprocessors (CMPs) either for energy efficiency or due to process variation.We propose Dynamic Core Boosting (DCB), a software-hardware cooperative system that mitigates the workload imbalance problem in performance asymmetric CMPs. Relying on dynamic voltage and frequency scaling to accelerate individual cores at a fine granularity, DCB attempts to balance the workloads by detecting and boosting critical threads. DCB coordinates its compiler and runtime to enable asymmetric CMPs to achieve near-optimal utilization of core boosting. The compiler instruments the program with instructions to give progress hints and the runtime monitors their execution, enabling DCB to intelligently accelerate selected threads within a total core boosting budget for better performance. On a simulated eight core system of varying frequency, our experiments using PARSEC benchmarks show that DCB improves the overall performance by an average of 33%, outperforming a reactive boosting scheme by an average of 10%.

Using Application-Level Thread Progress Information to Manage Power and Performance

“…The total number of iterations (the variable ntimes, passed to tb total) can be different for different threads and from one barrier-interval to the next. This style of software annotation is similar to previous work [10]- [12]. Instrumentation is a one-time effort.…”

“…Meeting Points [10] is an early work that uses application-level thread progress reporting annotation, but is only applicable to parallel loops with balanced workloads. Dynamic Core Boosting [12] is a recent work that uses application-level thread progress reporting annotation that accommodates imbalanced workloads, like ThreadBeats. However, the focus of that work is on accelerating cores in asymmetric chip multiprocessors using coarse-grained boosting techniques, whereas ThreadBeats is focused on tuning finegrained per-core DVFS in symmetric chip multiprocessors.…”

“…A more reliable way to measure the individual progress of different threads is to instrument them at the application level [10]- [12]. The application knows how much useful work a thread has actually completed and how much it has left to do before the next synchronization point.…”

CATA: Criticality Aware Task Acceleration for Multicore Processors

“…However, it only works for balanced programs, in which all threads run the same loop count with similar amount of computation. DCB [40] extends the model to unbalanced loops and to threads with different code in pipeline parallel programs, using an epoch-based approach in the latter case. However, it requires a significant profiling and instrumentation of the application.…”