« Previous month (Jun 2007) | Main | Next month (Aug 2007) »
Tuesday Jul 17, 2007

slog blog (or blogging on slogging)

slog_blog I've been slogging for a while on support for separate intent logs (slogs) for ZFS. Without slogs, an intent log is allocated dynamically from the main pool. It consists of a chain of varying block sizes which are anchored in fixed objects. Specifying a separate log device enables the use of limited capacity but fast block devices such NVRAM  and Solid State Drives (SSDs).

Using chained logs (clogs?) can also lead to pool fragmentation. This is because log blocks are allocated and then freed as soon as the pool transaction group has committed. So we get a swiss cheesing effect.

Interface

        zpool create <pool> <pool devices> log <log devices>

Creates a pool with separate intent log device(s). If more than one log device is specified then writes are load-balanced between devices. It's also possible to mirror log devices. For example a log consisting of two sets of two mirrors could be created thus:

                zpool create whirl <pool devices> \
log mirror c1t8d0 c1t9d0 \
mirror c1t10d0 c1t11d0

        zpool add <pool> log <log devices>
Creates a log device if it doesn't exist, or adds extra log devices if it does.

        zpool replace <pool> <old device> <new device>

Replace old log device with new log device.

        zpool attach <pool> <log device> <new log device>
Attaches a new log device to an existing log device. If the existing device is not a mirror then a 2 way mirror is created. If device is part of a two-way log mirror, attaching new_device creates a three-way log mirror, and so on.

        zpool detach pool <log device>

Detaches a log device from a mirror.

        zpool status

Additionally displays the log devices

        zpool iostat
Additionally shows IO statistics for log devices.

When a slog is full or if a non mirrored log device fails then ZFS will start using chained logs within the main pool.

Performance

The performance of databases and NFS is dictated by the latency of making data stable. They need to be assured that their transactions are not lost on power or system failure. So they are heavily dependent on the speed of the intent log devices.
Here's some database performance testing results:

  • Test program creates 32 threads and each does 8K O_DSYNC writes randomly to a 400MB byte file.
  • Test hardware was a Sun X4500 (aka thumper) with 48 x 500GB disks.
  • The NVRAM is the battery backed pci Micro Memory pci1332,5425 card.
  • Table values are MB/s

 Main pool disks

1
 2
 4
 8
 16
 32
0 slogs
 11
 14
 17
 15
 16
 13
1 slog
 12
 12
 12
 12
 12
 11
2 slogs
 17
 17
 17
 19
 19
 16
4 slogs
 17
 16
 15
 15
 16
 16
8 slogs
 18
 19
 20
 18
 16
 18
NVRAM
 221
 221
 218
 217
 215
 217


I also ran the same without write disk cache flushing
(echo zfs_nocacheflush/W 1 | mdb -kw)
Note, this should not be done on a real system unless the device cache is non-volatile.

Main pool disks

1
 2
 4
 8
 16
 32
0 slogs
 33
 83
 123
 136
 142
 143
1 slog
 45
 46
 44
 45
 45
 46
2 slogs
 97
 99
 90
 94
 94
 95
4 slogs
 124
 125
 127
 124
 127
 127
8 slogs
 135
 137
 134
 138
 138
 138
NVRAM
 225
 220
 226
 226
 226
 227

Note, these tables can be a bit mis-leading.  If you had 2 disks you'd have a choice of 2 main pool device  or 1 slog and 1 main pool device. So looking at the table  you should compare the following entries:
  • 2 main pool: 83MB/s
  • 1 slog, 1 main pool: 45MB/s
The first table highlights some issues of scaling which will be investigated further.

Perf summary

For this micro-benchmark and from limited other perf testing it makes sense to only use fast devices for the slog. However, there may be some cases where using regular disks as slog disks is faster than putting the same disks in the main pool.

Status/Bugs

This support was recently putback into Solaris Nevada build snv_68. Here's a list of slog bugs - fixed and to be fixed.

6574298  "slog still uses main pool for dmu_sync()"   - now fixed in snv_69
6574286 "removing a slog doesn't work"
6575965 "panic/thread=2a1016b5ca0: BAD TRAP: type=9 ...:" - panic when no main pool devices present - now fixed in snv_83


Posted at 03:23PM Jul 17, 2007 by perrin in ZFS  |  Comments[5]