Alita: Comprehensive Performance Isolation through Bias Resource Management for Public Clouds

摘要: The tenants of public cloud platforms share hard-ware resources on the same node, resulting in potential for performance interference (or malicious attacks). A tenant is able to degrade its neighbors node significantly through overuse shared memory bus, last level cache (LLC)/memory bandwidth, and power. To eliminate such unfairness we propose Alita, a runtime system consisting an online identifier adaptive eliminator. monitors hardware system-level event statistics identify resource polluters. eliminator improves normal applications by throttling only usage Specifically, Alita adopts bus lock sparsification, bias LLC/bandwidth isolation, selective power throttle Results experimental platform in-production with 30,000 nodes demonstrate that co-located virtual machines presence polluters based knowledge.