Deciphering Predictive Schedulers for Heterogeneous-ISA Multicore Architectures

Abstract

Heterogeneous architectures have become increasingly common. From co-packaging small and large cores, to GPUs alongside CPUs, to general-purpose heterogeneous-ISA architectures with cores implementing different ISAs. As diversity of execution cores grows, predictive models become of paramount importance for scheduling and resource allocation. In this paper, we investigate the capabilities of performance predictors in a heterogeneous-ISA setting, as well as the predictors' effects on scheduler quality. We follow an unbiased feature selection methodology to identify the optimal set of features for this task, instead of pre-selecting features before training. We propose metrics that bridge the gap between traditional prediction accuracy metrics and a scheduler's performance. We further present our evaluation methodology, which was meticulously designed with this study in mind, and finally, we incorporate our findings in MLbased schedulers and evaluate their sensitivity to the underlying system's level of heterogeneity.

Agon: A Scalable Competitive Scheduler for Large Heterogeneous Systems

“…Numerous proposals from the literature focus primarily on the first step of scheduling, i.e., workload-core performance prediction. Some simply assume the existence of a fast, optimal scheduler and focus entirely on improving prediction accuracy [16], [29]- [31], while others also attempt to alleviate its overhead. Two methods are commonly utilized to achieve that -(a) iterative methods perform job migrations upon scheduling intervals, slowly progressing towards a better schedule but likely sampling poor schedules along the way [31]- [33], and (b) divide-and-conquer methods that logically divide large systems into smaller independent scheduling islands [34], allowing the use of more complex algorithms operating with limited-scope knowledge.…”

“…Combined, hardware and software heterogeneity can significantly impact a scheduler's ability to find a good job assignment. Prior work proposed an Ease-of-Scheduling (EoS) metric that quantifies the impact of these factors in respect to the scheduler's outcome in a single number [29]. Figure 2 shows that for each system size (4/8/16 cores) scheduling difficulty can vary significantly, with some systems being quite easy to schedule, even for a random scheduler, while others can be quite challenging.…”

Byoc

“…Prodromou et al [46] tackle heterogeneous-ISA scheduling. They apply machine learning techniques to predict the ISA affinity of different application regions.…”

“…In the world of general-purpose compute, processor cores have emerged as reusable Intellectual Property (IP) blocks, with different, ISA-dependent features such as compact code density, low power operation, security extensions, high-throughput vector processing, and/or number of architectural registers. As a result of this, heterogeneous-ISA processors are emerging both as products [20] and as a hot topic of active research [16,46,[52][53][54].…”

Cross-ISA execution of SIMD regions for improved performance

“…Barbalace, et al 's [10] scheduler balances thread counts across ISA-different cores. Prodromou, et al [76] presents a machine learning-based program performance predictor that drives an ML-based heterogeneous-ISA job scheduler. These works have scoped out migration costs as well as SIMD workloads.…”