Platinum: A CPU-Efficient Concurrent Garbage Collector for Tail-Reduction of Interactive Services

The service-oriented architecture decomposes a monolithic service into single-purpose services for better modularity and reliability. The interactive nature, plus the fact of running inside a managed runtime, makes garbage collection a key to the reduction of tail latency of such services. However, prior concurrent garbage collectors reduce stopthe-world (STW) pauses by consuming more CPU resources, which can affect the application performance, especially under heavy workload.

Based on an in-depth analysis of representative latencysensitive workloads, this paper proposes Platinum, a new concurrent garbage collector to reduce the tail latency with moderate CPU consumption. The key idea is to construct an isolated execution environment for concurrent mutators to improve application latency without interfering with the execution of GC threads. Platinum further leverages a new hardware feature (i.e., memory protection keys) to eliminate software overhead in previous concurrent collectors. An evaluation against state-of-the-art concurrent garbage collectors shows that Platinum can significantly reduce the tail latency of real-world interactive services (by as much as 79.3%) while inducing moderate CPU consumption.