A New Family of Thread to Core Allocation Policies for an SMT ARM Processor

TL;DR

This paper introduces SYNPA, a new thread-to-core allocation policy for SMT ARM processors that uses a novel performance prediction model based on the ISC stack, significantly improving turnaround time.

Contribution

It proposes the ISC stack for ARM processors and develops SYNPA, a family of T2C policies that enhance performance prediction and allocation accuracy.

Findings

SYNPA4 reduces turnaround time by 38% over Linux.

The ISC stack improves performance prediction accuracy.

SYNPA outperforms existing policies by 3x on ARM processors.

Abstract

Modern high-performance servers commonly integrate Simultaneous Multithreading (SMT) processors, which efficiently boosts throughput over single-threaded cores. Optimizing performance in SMT processors faces challenges due to the inter-application interference within each SMT core. To mitigate the interference, thread-to-core (T2C) allocation policies play a pivotal role. State-of-the-art T2C policies work in two steps: i) building a per-application performance stack using performance counters and ii) building performance prediction models to identify the best pairs of applications to run on each core. This paper explores distinct ways to build the performance stack in ARM processors and introduces the Instructions and Stalls Cycles (ISC) stack, a novel approach to overcome ARM PMU limitations. The ISC stacks are used as inputs for a performance prediction model to estimate the applications' performance considering the inter-application interference. The accuracy of the prediction model (second step) depends on the accuracy of the performance stack (first step); thus, the higher the accuracy of the performance stack, the higher the potential performance gains obtained by the T2C allocation policy. This paper presents SYNPA as a family of T2C allocation policies. Experimental results show that $SYNPA4$, the best-performing SYNPA variant, outperforms turnaround time by 38\% over Linux, which represents 3$\times$ the gains achieved by the state-of-the-art policies for ARM processors. Furthermore, the multiple discussions and refinements presented throughout this paper can be applied to other SMT processors from distinct vendors and are aimed at helping performance analysts build performance stacks for accurate performance estimates in real processors.