1176504 - Consider increasing the size limit on the store buffer

Status: RESOLVED INVALID

(Core :: JavaScript: GC, defect)

Description

[Terrence Cole [:terrence]]

Now that we have telemetry on the minor GC reason, we can see that ~41% of minor collections are caused by running out of room in the store buffer.

This seems like a tradeoff between latency and throughput. If we have lots of cross generation edges, that means we'll preserve more of the nursery, which means we want to GC when the nursery is smaller for faster individual minor GC's. On the other hand, there are fixed overheads to any GC, like visiting the roots list, that are going to reduce throughput if we do them too much.

The actual performance tradeoff isn't super clear here however. I'll take a look at these numbers on Octane and see if I can find a good balance.

Comment 1

[Terrence Cole [:terrence]]

It doesn't actually work that way.