runtime: background scavenger is overzealous with small heaps #32012
Labels
Milestone
Comments
mknyszek
added
the
NeedsFix
|
Another thought occurred to me: this may obviate the need for heap-growth scavenging, since that case would now be covered by scavenging when allocating from scavenged memory. |
|
Change https://golang.org/cl/177097 mentions this issue: |
|
Change https://golang.org/cl/177237 mentions this issue: |
gopherbot
pushed a commit
that referenced
this issue
May 20, 2019
Currently, this test allocates many objects and relies on heap-growth scavenging to happen unconditionally on heap-growth. However with the new pacing system for the scavenging, this is no longer true and the test is flaky. So, this change overhauls TestPhysicalMemoryUtilization to check the same aspect of the runtime, but in a much more robust way. Firstly, it sets up a much more constrained scenario: only 5 objects are allocated total with a maximum worst-case (i.e. the test fails) memory footprint of about 16 MiB. The test is now aware that scavenging will only happen if the heap growth causes us to push way past our scavenge goal, which is based on the heap goal. So, it makes the holes in the test much bigger and the actual retained allocations much smaller to keep the heap goal at the heap's minimum size. It does this twice to create exactly two unscavenged holes. Because the ratio between the size of the "saved" objects and the "condemned" object is so small, two holes are sufficient to create a consistent test. Then, the test allocates one enormous object (the size of the 4 other objects allocated, combined) with the intent that heap-growth scavenging should kick in and scavenge the holes. The heap goal will rise after this object is allocated, so it's very important we do all the scavenging in a single allocation that exceeds the heap goal because otherwise the rising heap goal could foil our test. Finally, we check memory use relative to HeapAlloc as before. Since the runtime should scavenge the entirety of the remaining holes, theoretically there should be no more free and unscavenged memory. However due to other allocations that may happen during the test we may still see unscavenged memory, so we need to have some threshold. We keep the current 10% threshold which, while arbitrary, is very conservative and should easily account for any other allocations the test makes. Before, we also had to ensure the allocations we were making looked large relative to the size of a heap arena since newly-mapped memory was considered unscavenged, and so that could significantly skew the test. However, thanks to the fix for #32012 we were able to reduce memory use to 16 MiB in the worst case. Fixes #32010. Change-Id: Ia38130481e292f581da7fa3289c98c99dc5394ed Reviewed-on: https://go-review.googlesource.com/c/go/+/177237 Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Today the background scavenger will treat newly-mapped heap memory as unscavenged, which means that if you have an application with a small heap (such as 24 MiB) on Linux, then it will scavenge 64 - 24 = 40 MiB, which is a fair bit of work to be doing.
However, those 40 MiB were never touched in the first place, so they haven't been faulted in and don't count against your RSS anyway. The scavenging work is totally unnecessary.
Instead, we should probably treat newly-mapped memory as scavenged so that this issue is avoided. It also more accurately represents reality, and by tracking this in
HeapReleased, the runtime will have a better measure of the application's heap RSS.The one caveat here is that on Windows we currently map new heap memory as
MEM_COMMIT|MEM_RESERVEwhich means that Windows counts all of that against your RSS. But there's a clear solution here too: if we just map memory asMEM_RESERVEand then only map it asMEM_COMMITupon allocation, we can get this right.There isn't currently a noticeable performance impact from this so I don't think this is terribly high priority, but it could affect the "warm-up" time of applications on systems without huge pages and with larger arena sizes (e.g. Linux w/o THP, AIX, etc.). Aside from that, it would also be nice to just make the accounting a little bit better here for small heaps.
CC @aclements
The text was updated successfully, but these errors were encountered: