Closed Bug 1169116 Opened 9 years ago Closed 9 years ago

Trigger GC before measuring USS

Categories

(Firefox OS Graveyard :: Gaia::PerformanceTest, defect)

ARM
Gonk (Firefox OS)
defect
Not set
normal

Tracking

(Not tracked)

RESOLVED FIXED
FxOS-S3 (24Jul)

People

(Reporter: zbraniecki, Assigned: Eli)

References

Details

Attachments

(1 file)

It would give us more reliable metrics if we triggered GC before taking USS measurement in raptor.

So the order should be:

1) Record fullyLoaded timestamp
2) perform GC
3) Read USS

:gwagner suggest using http://mxr.mozilla.org/mozilla-central/ident?i=forceShrinkingGC
Depends on: 1169788
We've been tracking a memory regression that was reproducible on *some* devices for bug 1175260 for the last couple of weeks.

Yesterday, as an experiment I put fullyLoaded mark in setTimeout with 10s delay. Regression is gone.

It seems to me that without this patch measuring USS is basically measuring it at a random point wrt. GC and may be unreliable or platform dependent (depending on available memory Gecko may choose to GC at different points in time).
Depends on: 1169775
Depends on: 1177115
Depends on: 1177138
Assignee: nobody → eperelman
Target Milestone: --- → FxOS-S3 (24Jul)
Status: NEW → ASSIGNED
Comment on attachment 8632993 [details] [review]
Link to Github pull-request: https://github.com/eliperelman/mozilla-raptor/pull/2

LGTM with latest updates.

- verified --memoryDelay introduces delay after receiving fullyLoaded marker and before grabbing b2g-info memory values (coldlaunch on aries)
- gc is performed after each launch (fullyLoaded) and before grabbing memory; don't know how to verify an actual GC was done; so I am assuming the code is correct for that
Attachment #8632993 - Flags: review?(rwood) → review+
I've been using raptor with this patch and I did notice a drop in memory reports (from ~11.100 to ~10.600) which I can clearly attribute to the GC.

My stdev also dropped from ~100kb, to ~30kb.
Aries, DEBUG=raptor:* raptor test coldlaunch --app settings

Without patch:
| coldlaunch.uss                   | 19.100   | 19.100   | 19.100   | 19.100   | 0.000  | n/a |
| coldlaunch.pss                   | 23.400   | 23.400   | 23.400   | 23.400   | 0.000  | n/a |
| coldlaunch.rss                   | 38.100   | 38.100   | 38.100   | 38.100   | 0.000  | n/a |

With patch:
| coldlaunch.uss                   | 16.900   | 16.900   | 16.900   | 16.900   | 0.000  | n/a |
| coldlaunch.pss                   | 20.900   | 20.900   | 20.900   | 20.900   | 0.000  | n/a |
| coldlaunch.rss                   | 36.200   | 36.200   | 36.200   | 36.200   | 0.000  | n/a |


Noticeable difference, nice!
(In reply to Zibi Braniecki [:gandalf][:zibi] from comment #4)
> I've been using raptor with this patch and I did notice a drop in memory
> reports (from ~11.100 to ~10.600) which I can clearly attribute to the GC.
> 
> My stdev also dropped from ~100kb, to ~30kb.

That's great, thanks :gandalf!
In master: https://github.com/eliperelman/mozilla-raptor/commit/a2dfa590f4388bc0c53beddeb6c1cf0ebf56bc40
Status: ASSIGNED → RESOLVED
Closed: 9 years ago
Resolution: --- → FIXED
You need to log in before you can comment on or make changes to this bug.

Attachment

General

Creator:
Created:
Updated:
Size: