Mako: a low-pause, high-throughput evacuating collector for memory-disaggregated datacenters

Abstract

Resource disaggregation has gained much traction as an emerging datacenter architecture, as it improves resource utilization and simplifies hardware adoption. Under resource disaggregation, different types of resources (memory, CPUs, etc.) are disaggregated into dedicated servers connected by high-speed network fabrics. Memory disaggregation brings efficiency challenges to concurrent garbage collection (GC), which is widely used for latency-sensitive cloud applications, because GC and mutator threads simultaneously run and constantly compete for memory and swap resources. Mako is a new concurrent and distributed GC designed for memory-disaggregated environments. Key to Mako’s success is its ability to offload both tracing and evacuation onto memory servers and run these tasks concurrently when the CPU server executes mutator threads. A major challenge is how to let servers efficiently synchronize as they do not share memory. We tackle this challenge with a set of novel techniques centered around the heap indirection table (HIT), where entries provide one-hop indirection for heap pointers. Our evaluation shows that Mako achieves 12 ms at the 90th-percentile pause time and outperforms Shenandoah by an average of 3× in throughput.