Monday, December 19, 2016

Golang’s Real-Time GC in Theory and Practice

Will Sewell (via Hacker News):

In this blog post, we’ll look at Go’s garbage collector. We’ll see how it works (the tricolor algorithm), why it works (achieving such short GC pauses), and most importantly, whether it works (benchmarking these GC pauses, and comparing them with other languages).

[…]

The GC still has two stop-the-world phases: the initial stack scan for root objects, and a termination of the mark phase. Excitingly, this termination phase has recently been eliminated. We will discuss this optimization later. In practice we found the pause times of these phases to be <1ms with very large heaps.

[…]

The key takeaway from this investigation is that GCs are either optimized for lower latency or higher throughput. They might also perform better or worse at these depending on the heap usage of your program. (Are there a lot of objects? Do they have long or short lifetimes?)

Update (2016-12-20): Mike Hearn (Hacker News):

The reality is that Go’s GC does not really implement any new ideas or research. As their announcement admits, it is a straightforward concurrent mark/sweep collector based on ideas from the 1970s. It is notable only because it has been designed to optimise for pause times at the cost of absolutely every other desirable characteristic in a GC.

Comments RSS · Twitter

Leave a Comment