Age | Commit message (Collapse) | Author |
|
Looking at the dircache debug output, we see
that a "git status" workload has a very bad
cache hit rate because the entries expire or
get evicted before they can be reused.
Increase both cache size and lifetime for
a 4x speedup:
Before: 75s
After: 17s
https://github.com/rfjakob/gocryptfs/issues/410
|
|
Before:
Lookup "errno.html/1/2/3/4/5": miss
Store: "errno.html/1/2/3/4/5" fd=26 iv=21be6e083d60dcabfe7368264d5082b7
Lookup "errno.html": hit 25 6d68a16d217978915036a3bd55428ae7
Lookup "errno.html/1": hit 25 932a464c299b3430c5e55c924f98ac4d
Lookup "errno.html/1/2": hit 25 7d53348b1692d537f017bf86b3cf5feb
Lookup "errno.html/1/2/3": hit 25 2aef1c9d1ab2b55b163215053fefe703
Lookup "errno.html/1/2/3/4": hit 25 cb802be53721c46a46247c5e4e0f4ce6
Lookup "errno.html/1/2/3/4": hit 25 cb802be53721c46a46247c5e4e0f4ce6
Lookup "errno.html": hit 25 6d68a16d217978915036a3bd55428ae7
After:
Lookup "earlyoom/.git/refs" hit fd=10 dup=17 iv=6ae2cecd269a25e8d946aff6afe9b8b8
Lookup "earlyoom/.git/refs/remotes" hit fd=19 dup=17 iv=f04c2d2a5bcc33ebdeaca664859c980d
Lookup "earlyoom/.git/refs/remotes/origin" miss
Store "earlyoom/.git/refs/remotes/origin" fd=17 iv=834a64a1697c9f5705455ba6dbed22b5
Lookup "earlyoom" hit fd=7 dup=25 iv=2303a892d6e2357c483574a8070b7679
Lookup "earlyoom/.git" hit fd=11 dup=25 iv=d43ca4aff23720c57789c9f62f0aee00
Lookup "earlyoom/.git" hit fd=11 dup=25 iv=d43ca4aff23720c57789c9f62f0aee00
Lookup "earlyoom/.git/refs" hit fd=10 dup=25 iv=6ae2cecd269a25e8d946aff6afe9b8b8
Lookup "earlyoom/.git/refs/heads" hit fd=13 dup=25 iv=f9245e7c066b9adc768a1a666da9fbc8
|
|
The files are apparently processed in alphabetic order, so cli_args.go is
processed before main.go. In order to run before the go-fuse imports, put
the 'ensure fds' code in a separate package. Debug messages are omitted
to avoid additional imports (that might contain other code messing up our
file descriptors).
|
|
|
|
As the dirCache now has 3 entries, the tests should accept
up to 3 extra fds without declaring an fd leak.
|
|
The missing break meant that we may find a second
hit in the cache, Dup() a second fd, and leak the first
one.
Thanks @slackner for finding this.
|
|
This was inadvertedly kept enabled after benchmarking.
|
|
3 entries should work well for up to three parallel users.
It works well for extractloop.bash (two parallel tar extracts).
|
|
|