2021/02/18 by Bodhisatwa Chatterjee, Sharjeel Khan, Chatterjee, Bodhisatwa +3
Computer Science · #Advanced Data Storage Technologies #Cloud Computing and Resource Management #FOS: Computer and information sciences #FOS: Electrical engineering #Hardware Architecture (cs.AR) #Parallel Computing and Optimization Techniques #Programming Languages (cs.PL) #Systems and Control (eess.SY) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2102.09673
openalex publication_date 2021/02/18 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
With a growing number of cores in modern high-performance servers, effective\nsharing of the last level cache (LLC) is more critical than ever. The primary\nagenda of such systems is to maximize performance by efficiently supporting\nmulti-tenancy of diverse workloads. However, this could be particularly\nchallenging to achieve in practice, because modern workloads exhibit dynamic\nphase behaviour, which causes their cache requirements & sensitivities to vary\nat finer granularities during execution. Unfortunately, existing systems are\noblivious to the application phase behavior, and are unable to detect and react\nquickly enough to these rapidly changing cache requirements, often incurring\nsignificant performance degradation. In this paper, we propose Com-CAS, a new\napportioning system that provides dynamic cache allocations for co-executing\napplications. Com-CAS differs from the existing cache partitioning systems by\nadapting to the dynamic cache requirements of applications just-in-time, as\nopposed to reacting, without any hardware modifications. The front-end of\nCom-CAS consists of compiler-analysis equipped with machine learning mechanisms\nto predict cache requirements, while the back-end consists of proactive\nscheduler that dynamically apportions LLC amongst co-executing applications\nleveraging Intel Cache Allocation Technology (CAT). Com-CAS's partitioning\nscheme utilizes the compiler-generated information across finer granularities\nto predict the rapidly changing dynamic application behaviors, while\nsimultaneously maintaining data locality. Our experiments show that Com-CAS\nimproves average weighted throughput by 15% over unpartitioned cache system,\nand outperforms state-of-the-art partitioning system KPart by 20%, while\nmaintaining the worst individual application completion time degradation to\nmeet various Service-Level Agreement (SLA) requirements.\n