1262321 - Poor GC throughput in long-running session persistence task

Status: RESOLVED FIXED

(Core :: JavaScript: GC, defect)

Description

[The 8472]

Attachment: ffgc.7z

This is a followup to bug 1260475. I have captured GC logs over the uptime of several hours and looked at a few long jank periods with the profiler. It turns out that most of it is caused by session persistence (bug 912717) and the sessionmanager addon (filed a bug on their tracker).

Nevertheless both of those root causes *also* cause a massive amount of GC as knock-on effect.

Attached the compressed GC log.

Comment 1

[The 8472]

Attachment: GC-utilization.png

The screenshot of the profiler shows a time span where the addon creates a session backup. This takes about 30 of wall time.

Based on the matched method name it seems that it actually spends its majority of time in GC. In other words, poor collector performance is responsible for the majority of the jank.

Comment 2

[The 8472]

Attachment: inverse call graph.png

The inverse call graph for a similar event shows drainMarkStack/processMarkStack as the top item.

So I guess bug 638660 would be the lowest-hanging fruit here in terms of walltime throughput, although not in CPU cycles throughput.

Comment 3

[The 8472]

I'm no longer experiencing this issue.