PInTE: Probabilistic Induction of Theft Evictions

Cesar Gomes
Xuesi Chen
Mark Hempstead

IEEE International Symposium on Workload Characterization (IISWC)
Austin, Texas, November 2022

Abstract

Cache contention analysis remains complex without a controlled & lightweight method of inducing contention for shared resources. Prior art commonly leverages a second workload on an adjacent core to cause contention, and the workload is either real or tune-able. Using a secondary workload comes with unique problems in simulation: real workloads aren’t controllable and can result in many combinations to measure a broad range of contention; and tune-able workloads provide control but don’t guarantee contention without filling all cache sets with contention behavior. Lastly, running multiple workloads increases the run-time of simulation environments by 2.4X on average.

We introduce Probabilistic Induction of Theft Evictions, or PInTE which allows controllable contention induction via data movement towards eviction in the last level cache replacement policy. PInTE provides configurable contention with 2.6X fewer experiments, 2.2X less average time, and 5.6$X less total time for a set of SPEC 17 speed-based traces. Further, PInTE incurs -8.46% average relative error in performance when compared to real contention. Run-time and reuse behavior of workloads under PInTE contention approximate behavior under real contention — information distance is 0.03 bits and 0.84 bits, respectively. Additionally, PInTE enables a first-time contention sensitivity analysis of SPEC and case studies which evaluate the resilience of micro-architectural techniques under growing contention.