diff options
| author | Misha Wagner <mishaw@google.com> | 2021-09-20 13:53:00 +0100 |
|---|---|---|
| committer | Misha Wagner <mishaw@google.com> | 2021-09-20 18:16:56 +0000 |
| commit | 665e91e17196c499462553e2c463c39727ba8738 (patch) | |
| tree | e11b4ca258bfd7956588ef4e15b993fa3f81e854 /rs/java/android/renderscript/ProgramFragmentFixedFunction.java | |
| parent | 7d611c750c3d518794b2e580539bf9a22222a41f (diff) | |
Use Process.getRss instead of mProfile.getLastRss in CacheOomRanker.
RSS values are cached in ProcessStateRecord, and repopulated when they
are >10s old.
mProfile.getLastRss values can be 15s-2h stale, which is why we're using
this implementation. My understanding is that the best solution would be
to populate mProfile.getLastRss more often - but it is currently
populated alongside PSS, which is more expensive to calculate.
Therefore, rather than the more complex change of splitting up PSS & RSS
collection, we opt for this simple (if not optimal) approach as
CacheOomRanker is still experimental. Upon successful launch, we can
revisit this.
When logging the duration of CacheOomRanker calculations, we saw no
significant increase in computation time with this change. If we turn
off caching RSS values, we see a 10x increase in computation time.
Test: atest CacheOomRankerTest
Bug: 196031723
Merged-In: I8e80db2aab338df07f07842131a9d4b55bebd054
Change-Id: I8e80db2aab338df07f07842131a9d4b55bebd054
(cherry picked from commit 8bd226e360e0594f3421fa12abfa36340df170e4)
Diffstat (limited to 'rs/java/android/renderscript/ProgramFragmentFixedFunction.java')
0 files changed, 0 insertions, 0 deletions