This repository has been archived by the owner on Oct 30, 2018. It is now read-only.
Reduce maximum memory usage via tempfiles #191
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
added 2 commits
December 6, 2017 15:54
Following https://golang.org/pkg/runtime/pprof/. When attempting to build images in https://github.com/linuxkit/kubernetes CI the process is mysteriously being SIGKILL'd, which I think might be down to OOMing due to the resource limits placed on the build container. I haven't done so yet but I'm intending to use these options to investigate and they seem potentially useful in any case, even if this turns out to be a red-herring. Signed-off-by: Ian Campbell <ijc@docker.com>
All of the `output*` functions took a `[]byte` and immediately wrapped it in a `bytes.Buffer` to produce an `io.Reader`. Make them take an `io.Reader` instead and satisfy this further up the call chain by directing `moby.Build` to output to a temp file instead of another `bytes.Buffer`. In my test case (building kube master image) this reduces Maximum RSS (as measured by time(1)) from 6.7G to 2.8G and overall allocations from 9.7G to 5.3G. When building a tar (output to /dev/null) the Maximum RSS fell slightly from 2.2G to 2.1G. Overall allocations remained stable at around 5.3G. Signed-off-by: Ian Campbell <ijc@docker.com>
Rather than queueing up into a `bytes.Buffer`. In my test case (building kube master image) this reduces Maximum RSS (as measured by time(1)) compared with the previous patch from 2.8G to 110M. The tar output case goes from 2.1G to 110M also. Overall allocations are ~715M in both cases. Signed-off-by: Ian Campbell <ijc@docker.com>
ijc
force-pushed
the
reduce-memory-via-tempfiles
branch
from
c3497b9
to
3045a80
Compare
|
FWIW the only remaining |
justincormack
approved these changes
Dec 12, 2017
ijc
pushed a commit
to ijc/moby-tool
that referenced
this pull request
Dec 13, 2017
This was introduced by moby#191 but somehow did not trigger either for me in local testing or in CI. It did trigger in initial CI of linuxkit/linuxkit#2811 which can be seen at https://linuxkit.datakit.ci/linuxkit/linuxkit/pr/2811?history=1637690296123e9a15307b3a41b290da6e27e7cc The error is: Failed to docker rm container «...»: «...»: aufs: unmount error after retries: «...»: device or resource busy No doubt because we were still holding an open fd while trying to remove the container. Unclear why this didn't repro for me (docker 17.11.0-ce with overlay2) or whatever CI uses. Signed-off-by: Ian Campbell <ijc@docker.com>
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Replacing a few uses of
bytes.Bufferwith temporary files or passing aroundio.Readerinstead of a full buffer reduces the Maximum RSS (measured bytime(1)) from 6.7G to 110M when building thekube-master.isofrom https://github.com/linuxkit/kubernetes.Motivation for looking into this was that in the https://github.com/linuxkit/kubernetes CI the image build process is occasionally being SIGKILL'd, since the CI containers are limited to 4G I hypothesised that the process was being OOM killed (I have not yet tested with these changes in place).
Since i have plenty of RAM no my test system the data, presumably, remains hot in the page cache so elapsed wallclock time is unaffected (it's a few seconds faster on a minute long build, so in the noise). On a memory (but not disk) constrained system it ought to complete slowly instead of getting killed.
Also adds
-memprofileand-cpuprofileoptions to enable measurement, although I mostly ended up using `time(1) since the profiles show cumulative or overall allocations rather than max rss.I'm not sure but with these changes it might be possible to mark some additional image formats as streaming.