1279396 - Tp5o Private Bytes should be split/fixed

Status: RESOLVED FIXED

(Testing :: Talos, defect)

Description

[Mike Hommey [:glandium]]

I was looking at talos results for current jemalloc dev branch vs. jemalloc 4.2, and saw a jump in Private Bytes. But while looking at the numbers, I also saw that the value itself made no sense, because it was also an order of magnitude larger than RSS, which according to what Private Bytes usually means, can't happen.

Erahm pointed me to https://dxr.mozilla.org/mozilla-central/source/testing/talos/talos/cmanager_linux.py#75-102 which indicates that what this actually measures is MAP_PRIVATE address space allocated by the process. Which, in an of itself, *is* an interesting metric. But:
- that doesn't match any definition of Private Bytes
- that doesn't match what Private Bytes measure on other platforms

So I think we should have measure for address space consumption *and* a more accurate measure of Private Bytes.

That being said, that also kind of(?) overlaps with AWSY...

Comment 1

[Eric Rahm [:erahm]]

This is a problem on Windows too; it's possible we already have a bug for it. There was discussion previously about just using psutil for USS, it's in-tree and I wrote the USS additions and AFAICT they are correct.

Mike, what would you suggest calling this measure (what we currently do on Linux)? Do you think we can get a similar measure for Windows?

Comment 2

[Joel Maher ( :jmaher ) (UTC -8)]

technically we do overlap with AWSY, but AWSY doesn't have the full coverage of what Talos does.  I would like to see us fix this sooner rather than later.

related is bug 1250169.

Comment 3

[Joel Maher ( :jmaher ) (UTC -8)]

I would like to resolve this sooner than later- I think all memory on win7, win10, linux64 should be done on AWSY- I would like to confirm that is the case and then we just have talos measuring memory on osx.

Comment 4

[Joel Maher ( :jmaher ) (UTC -8)]

we resolved this the other week