On Thu, Apr 19, 2012 at 11:46:02AM -0400, Josef 'Jeff' Sipek wrote:
> On Thu, Apr 19, 2012 at 09:48:21AM +1000, Dave Chinner wrote:
> > On Wed, Apr 18, 2012 at 11:11:40AM -0400, Josef 'Jeff' Sipek wrote:
> > > Greetings! I have a file server that get a pretty nasty load (about 15
> > > million files created every day). After some time, I noticed that the
> > > load
> > > average spiked up from the usual 30 to about 180. dmesg revealed:
> > >
> > > [434042.318401] INFO: task php:2185 blocked for more than 120 seconds.
> > > [434042.318403] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs"
> > > disables this message.
> > > [434042.318405] php D 000000010675d6cd 0 2185 27306
> > > 0x00000000
> > > [434042.318408] ffff88008d735a48 0000000000000086 ffff88008d735938
> > > ffffffff00000000
> > > [434042.318412] ffff88008d734010 ffff88000e28e340 0000000000012000
> > > ffff88008d735fd8
> > > [434042.318416] ffff88008d735fd8 0000000000012000 ffff8807ef9966c0
> > > ffff88000e28e340
> > > [434042.318419] Call Trace:
> > > [434042.318442] [<ffffffffa0087a9b>] ? xfs_trans_brelse+0xee/0xf7 [xfs]
> > > [434042.318464] [<ffffffffa00689de>] ? xfs_da_brelse+0x71/0x96 [xfs]
> > > [434042.318485] [<ffffffffa006df10>] ?
> > > xfs_dir2_leaf_lookup_int+0x211/0x225 [xfs]
> > > [434042.318489] [<ffffffff8141481e>] schedule+0x55/0x57
> > > [434042.318512] [<ffffffffa0083de2>] xlog_reserveq_wait+0x115/0x1c0 [xfs]
> > > [434042.318515] [<ffffffff810381f1>] ? try_to_wake_up+0x23d/0x23d
> > > [434042.318539] [<ffffffffa0083f45>] xlog_grant_log_space+0xb8/0x1be
> > > [xfs]
> > > [434042.318562] [<ffffffffa0084164>] xfs_log_reserve+0x119/0x133 [xfs]
> > > [434042.318585] [<ffffffffa0080cf1>] xfs_trans_reserve+0xca/0x199 [xfs]
> > > [434042.318605] [<ffffffffa00500dc>] xfs_create+0x18d/0x467 [xfs]
> > > [434042.318623] [<ffffffffa00485be>] xfs_vn_mknod+0xa0/0xf9 [xfs]
> > > [434042.318640] [<ffffffffa0048632>] xfs_vn_create+0xb/0xd [xfs]
> > > [434042.318644] [<ffffffff810f0c5d>] vfs_create+0x6e/0x9e
> > > [434042.318647] [<ffffffff810f1c5e>] do_last+0x302/0x642
> > > [434042.318651] [<ffffffff810f2068>] path_openat+0xca/0x344
> > > [434042.318654] [<ffffffff810f23d1>] do_filp_open+0x38/0x87
> > > [434042.318658] [<ffffffff810fb22e>] ? alloc_fd+0x76/0x11e
> > > [434042.318661] [<ffffffff810e40b1>] do_sys_open+0x10b/0x1a4
> > > [434042.318664] [<ffffffff810e4173>] sys_open+0x1b/0x1d
> > >
> > > It makes sense that'd the load average would spike up if some major lock
> > > got
> > > held longer than it should have been.
> > That's possibly just IO load. It's waiting for log space to become
> > available, and that will only occur when metadata is written back to
> > disk.
> Yeah, it could be...
> > What's the storage subsystem, what IO scheduler, the actually
> > workload that is running at the time the spike occurs (e.g did
> > someone run a directory traversal that changed every indoe?), and so
> > on.
> fs3.ess ~ # MegaCli -CfgDsply -aAll
> Adapter: 0
> Product Name: LSI MegaRAID SAS 9260-16i
> Memory: 512MB
> BBU: Present
> RAID Level : Primary-6, Secondary-0, RAID Level Qualifier-3
> /dev/sda4 /var/data/disk0 xfs
> 0 0
Only unusual thing is the filestreams allocator is in use...
> fs3.ess ~ # xfs_info /var/data/disk0/
> meta-data=/dev/sda4 isize=256 agcount=32, agsize=268435455
> = sectsz=512 attr=2
> data = bsize=4096 blocks=8519632640, imaxpct=5
> = sunit=0 swidth=0 blks
> naming =version 2 bsize=4096 ascii-ci=0
> log =internal bsize=4096 blocks=521728, version=2
and it is a large log, so that's a definite candidate for long
stalls pushing the tail of the log. The typical push that you'll get
stalled on is trying to keep 25% fthe log space free, so there's
potentially hundreds of megabytes of random 4/8k metadata writes to
be done to free that space in the log...
> = sectsz=512 sunit=0 blks, lazy-count=1
> realtime =none extsz=4096 blocks=0, rtextents=0
> As far as the workload is concerned, there's the typical (for us) load of
> create 15 million files in a day (pretty standard diurnal distribution as far
> as the load is concerned). Half the files are about 300 bytes, and the other
> half averages 20kB in size. On this system, the files are never read back.
> Since we had a fs corruption recently (due to a power failure & disk caches
> being on), our software managed to dump about 25 million files onto /. I've
> been rsync'ing them to the data partition at the same time (yes, it's slow
> because they are two partitions on the same array). The rsync aside, we've
> this workload going for a month on this particular server without any issues.
> (The other server which is set up identically has been fine too.) On Monday,
> we mkfs'd this server's sda4, mounted it as before and started the workload.
What was the size of the log on the previous incarnation of the
> Then, on Tuesday and Wednesday, we saw two loadavg spikes for no apparent
> reason. (Our system forks a bunch and then each child does network & disk
> so 180 loadavg makes sense if the system gums up.)
> The following applies to all of the spikes, but I'm specifically talking about
> the spike from this morning. During the ~45 minute spike, there seems to be
> very little disk I/O (<1 MByte/s compared to the usual 10 MBytes/s). Since
That sounds like it might have dropped into random 4k write IO or
inode cluster RMW cycles - a single large RAID6 volume is going to
be no faster than a single spindle at this. Can you get `iostat -d
-m -x 5` output when the next slowdown occurs so we can see the IOPS
and utilisation as well as the bandwidth?
> all the I/O children get stuck waiting for disk (ps says they are in 'D'
> state), the network utilization drops to essentially 0. The CPU usage also
> drops to ~0%.
> I *think* things return to normal because our code eventually notices that
> the children are stuck and it 'kill -9's them. (It's nicer at first.)
you can't kill processes stuck waiting for log space IIRC.
> > Also, getting the output of 'echo w > /proc/sysrq-trigger' whould be
> > helpful here....
> I assume you want this when the loadavg spikes up to 180. /me waits for the
> next time it happens.