Camelia, the Perl 6 bug

IRC log for #gluster-dev, 2012-10-17

| Channels | #gluster-dev index | Today | | Search | Google Search | Plain-Text | summary

All times shown according to UTC.

Time Nick Message
00:49 bala joined #gluster-dev
01:21 a2 joined #gluster-dev
01:21 _Bryan_ joined #gluster-dev
01:21 plantain joined #gluster-dev
04:20 sripathi joined #gluster-dev
04:38 deepakcs joined #gluster-dev
04:44 vpshastry joined #gluster-dev
05:15 puebele joined #gluster-dev
05:20 sgowda joined #gluster-dev
05:26 kshlm joined #gluster-dev
05:26 kshlm joined #gluster-dev
05:31 bala1 joined #gluster-dev
05:35 raghu joined #gluster-dev
06:15 hagarth joined #gluster-dev
06:27 kshlm joined #gluster-dev
06:27 kshlm joined #gluster-dev
06:40 kshlm joined #gluster-dev
06:40 kshlm joined #gluster-dev
06:48 deepakcs joined #gluster-dev
06:50 deepakcs joined #gluster-dev
07:08 lkoranda joined #gluster-dev
07:41 kshlm joined #gluster-dev
07:41 kshlm joined #gluster-dev
08:19 badone_home joined #gluster-dev
08:57 sgowda joined #gluster-dev
09:07 sripathi joined #gluster-dev
09:08 the-me johnmark: first I ask myself why it is required? except when we are in freeze time I will provide the most actual version of glusterfs always on the offical Debian mirrors
09:33 kshlm joined #gluster-dev
09:33 kshlm joined #gluster-dev
10:07 the-me joined #gluster-dev
10:17 sgowda joined #gluster-dev
10:21 sripathi joined #gluster-dev
10:21 sripathi joined #gluster-dev
10:27 lkoranda_ joined #gluster-dev
11:00 hagarth joined #gluster-dev
11:01 lkoranda joined #gluster-dev
11:58 gbrand_ joined #gluster-dev
12:03 gbrand_ Hello, is there a limit to the amount of ext. metadata that one can write per file? I am writing several extended metadata entries per file, but at a certain moment I get "No space left on device" from client3_3_setxattr_cbk. There is enough space and inodes available.
12:25 vpshastry left #gluster-dev
13:07 gbrand_ joined #gluster-dev
13:07 hagarth joined #gluster-dev
13:19 johnmark gbrand_: depends on the disk filesystem and options used when creating it
13:20 jdarcy I thought that both ext4 and XFS would expand into extension blocks instead of returning ENOSPC.  Could be wrong, or it could be a different FS.
13:32 kkeithley joined #gluster-dev
13:44 jdarcy I think I've figured out why we're still issuing tons of getxattrs despite md-cache.
13:44 jdarcy It's because md-cache doesn't fail cache misses even when the cache is considered valid.
13:46 sripathi joined #gluster-dev
13:50 hagarth joined #gluster-dev
14:05 gbrand_ joined #gluster-dev
14:07 gbrand_ joined #gluster-dev
14:31 wushudoin joined #gluster-dev
14:34 gbrand_ joined #gluster-dev
14:44 lkoranda_ joined #gluster-dev
14:50 gbrand__ joined #gluster-dev
15:04 lkoranda joined #gluster-dev
15:07 wushudoin joined #gluster-dev
15:17 lkoranda joined #gluster-dev
15:22 johnmark hagarth: ping
15:26 wushudoin joined #gluster-dev
15:28 semiosis the-me: people want to be able to install the latest glusterfs (currently 3.3.1) on their debian stable/squeeze systems withouth having to install packages from debian unstable or experiemntal
15:28 edward1 joined #gluster-dev
15:45 bala joined #gluster-dev
16:08 hagarth johnmark: pong
16:20 bala joined #gluster-dev
17:34 the-me semiosis: johnmark: this will be the case again if the freeze is over
17:34 gbrand_ joined #gluster-dev
18:14 a2 jdarcy, you mean -ve xattr caching in md-cache?
18:16 jdarcy a2: Yep.  Already pushed a simple version, just for fun.
18:16 jdarcy a2: Turns out that's only part of the problem, but still another place we can avoid wasted effort.
18:17 a2 what's the rest of the problem?
18:18 jdarcy a2: In some of my tests, I was seeing ~3ms per getdents call (translates into readdirp for us) that fetches 38 entries, but spaced 9ms apart.  That other 6ms is something other than network latency, probably ping-ponging through FUSE.
18:19 jdarcy BTW, the 38 entries was an improvement over the usual 23 because we know we don't have to include the size of an iatt against the request size, because we'll only be returning the dirent through that interface.
18:21 jdarcy It just really bugs me that directory listings seem to be incurably slow short of pushing entry/stat/xattr caching/prefetch into the kernel.
18:22 a2 jdarcy, the FUSE readdirplus patch fixed exactly that.. waiting for it to trickle down to a downstream kernel
18:23 a2 though xattrs are not a part of it.. so md-cache doing a good job will be very important
18:24 a2 http://git.kernel.org/?p=linux/kernel​/git/mszeredi/fuse.git;a=commit;h=2ca​d95b9c617d28d408adb7444ae3d40f4a6eff2 the commit log has some crude test results
18:24 jdarcy Unfortunately I think even if md-cache does a good job it still costs us a context switch per getxattr to get there.
18:25 jdarcy Maybe implement readdirplusplus to include essential xattrs?  ;)
18:26 a2 that would be cool
18:26 a2 unfortunately there is no xattr caching framework in fuse kernel module to easily leverage
18:26 a2 the attr/entry caching frameowrk existed, so readdirplus was easy by utilizing just them
18:27 jdarcy True, it'd be a significant piece of work.  Might belong *somewhere* on the road map, though.
18:27 a2 btw, with a real network the saving ratio must be much higher than the results in the commit log.. those tests were within the same physical machine between guest and host
18:54 wushudoin| joined #gluster-dev
19:08 hagarth1 joined #gluster-dev

| Channels | #gluster-dev index | Today | | Search | Google Search | Plain-Text | summary