Blade: A Data Center Garbage Collector

Abstract

An increasing number of high-performance distributed systems are written in garbage collected languages. This removes a large class of harmful bugs from these systems. However, it also introduces high tail-latency do to garbage collection pause times. We address this problem through a new technique of garbage collection avoidance which we call Blade. Blade is an API between the collector and application developer that allows developers to leverage existing failure recovery mechanisms in distributed systems to coordinate collection and bound the latency impact. We describe Blade and implement it for the Go programming language. We also investigate two different systems that utilize Blade, a HTTP load-balancer and the Raft consensus algorithm. For the load-balancer, we eliminate any latency introduced by the garbage collector, for Raft, we bound the latency impact to a single network round-trip, (48 μs in our setup). In both cases, latency at the tail using Blade is up to three orders of magnitude better.