Ginseng: market-driven LLC allocation

TLDR: Experimental results show that Ginseng for cache allocation improved clients' aggregated benefit by up to 42.8× compared with state-of-the-art static and dynamic algorithms.

Abstract: Cloud providers must dynamically allocate their physical resources to the right client to maximize the benefit that they can get out of given hardware. Cache Allocation Technology (CAT) makes it possible for the provider to allocate last level cache to virtual machines to prevent cache pollution. The provider can also allocate the cache to optimize client benefit. But how should it optimize client benefit, when it does not even know what the client plans to do? We present an auction-based mechanism that dynamically allocates cache while optimizing client benefit and improving hardware utilization. We evaluate our mechanism on benchmarks from the Phoronix Test Suite. Experimental results show that Ginseng for cache allocation improved clients' aggregated benefit by up to 42.8× compared with state-of-the-art static and dynamic algorithms.