Damien Fleuriot
2011-Feb-24 07:30 UTC
ZFS - abysmal performance with samba since upgrade to 8.2-RELEASE
Hello list,
I've recently upgraded my home box from 8.2-PRE to 8.2-RELEASE and since
then I've been experiencing *abysmal* performance with samba.
We're talking transfer rates of say 50kbytes/s here, and I'm the only
client on the box.
I've seen here and there discussions about sendfile's support, and how
it is recommended to disable it in samba.
Now, the man tells us:
Default: use sendfile = false
On the other side, FTP and SFTP transfers are OK, at 20mbytes/s or so.
Find below a bit of info
smb.conf
---
[global]
workgroup = MYGROUP
server string = Samba Server
security = user
hosts allow = 192.168.0.1 192.168.1.1 fe80::1
load printers = no
log file = /var/log/samba/log.%m
max log size = 50
# You may want to add the following on a Linux system:
socket options = SO_RCVBUF=8192 SO_SNDBUF=8192
dns proxy = no
bind interfaces only = yes
interfaces = re0 lagg0
hide unreadable = yes
[homes]
comment = Home Directories
browseable = no
writable = yes
loader.conf
---
# Tune ZFS somewhat aye ?
vm.kmem_size="3072M"
vm.kmem_size_max="3072M"
vfs.zfs.arc_min="128M"
vfs.zfs.arc_max="2048M"
vfs.zfs.txg.synctime="1"
vfs.zfs.prefetch_disable="0"
vfs.zfs.vdev.min_pending="4"
vfs.zfs.vdev.max_pending="8"
ZFS pool
---
nas# zpool status
pool: rtank
state: ONLINE
status: The pool is formatted using an older on-disk format. The pool can
still be used, but some features are unavailable.
action: Upgrade the pool using 'zpool upgrade'. Once this is done, the
pool will no longer be accessible on older software versions.
scrub: none requested
config:
NAME STATE READ WRITE CKSUM
rtank ONLINE 0 0 0
raidz2 ONLINE 0 0 0
gpt/zfs-ada0 ONLINE 0 0 0
gpt/zfs-ada1 ONLINE 0 0 0
gpt/zfs-ada2 ONLINE 0 0 0
gpt/zfs-ada3 ONLINE 0 0 0
gpt/zfs-ad14 ONLINE 0 0 0
gpt/zfs-ad10 ONLINE 0 0 0
gpt/zfs-ad11 ONLINE 0 0 0
gpt/zfs-ad12 ONLINE 0 0 0
gpt/zfs-ad13 ONLINE 0 0 0
zpool version 14
zfs version 3
So mmmm, sendfile disabled, ZFS pool not quite updated, good FTP/ssh
performance, horribad samba performance.
Does anyone have an idea of stuff I should be looking at ?
Again, this has only begun since I've upgraded from 8.2-PRE/RC3 to
8.2-RELEASE.
Jeremy Chadwick
2011-Feb-24 07:55 UTC
ZFS - abysmal performance with samba since upgrade to 8.2-RELEASE
On Thu, Feb 24, 2011 at 08:30:17AM +0100, Damien Fleuriot wrote:> Hello list, > > I've recently upgraded my home box from 8.2-PRE to 8.2-RELEASE and since > then I've been experiencing *abysmal* performance with samba. > > We're talking transfer rates of say 50kbytes/s here, and I'm the only > client on the box.I have a similar system with significantly less disks (two pools, one disk each; yes, no redundancy). The system can push, via SMB/CIFS across the network about 65-70MBytes/sec, and 80-90MByte/sec via FTP. I'll share with you my tunings for Samba, ZFS, and the system. I spent quite some time messing with different values in Samba and FreeBSD to find out what got me the "best" performance without destroying the system horribly. Please note the amount of memory matters greatly here, so don't go blindly setting these if your system has some absurdly small amount of physical RAM installed. Before getting into what my system has, I also want to make clear that there have been cases in the past where people were seeing abysmal performance from ZFS, only to find out it was a *single disk* in their pool which was causing all of the problems (meaning a single disk was performing horribly, impacting everything). I can try to find the mailing list post, but I believe the user offlined the disk (and later replaced it) and everything was fast again. Just a FYI. System specifications ======================* Case - Supermicro SC733T-645B * MB - Supermicro X7SBA * CPU - Intel Core 2 Duo E8400 * RAM - CT2KIT25672AA800, 4GB ECC * RAM - CT2KIT25672AA80E, 4GB ECC * Disk - Intel X25-V SSD (ada0, boot) * Disk - WD1002FAEX (ada1, ZFS "data" pool) * Disk - WD2001FASS (ada2, ZFS "backups" pool) Samba ======================Rebuild the port (ports/net/samba35) with AIO_SUPPORT enabled. To use AIO you will need to load the aio.ko kernel module (kldload aio) first. Relevant smb.conf tunings: [global] socket options = TCP_NODELAY SO_SNDBUF=131072 SO_RCVBUF=131072 use sendfile = no min receivefile size = 16384 aio read size = 16384 aio write size = 16384 aio write behind = yes ZFS pools ====================== pool: backups state: ONLINE scrub: none requested config: NAME STATE READ WRITE CKSUM backups ONLINE 0 0 0 ada2 ONLINE 0 0 0 errors: No known data errors pool: data state: ONLINE scrub: none requested config: NAME STATE READ WRITE CKSUM data ONLINE 0 0 0 ada1 ONLINE 0 0 0 errors: No known data errors ZFS tunings ======================Your tunings here are "wild" (meaning all over the place). Your use of vfs.zfs.txg.synctime="1" is probably hurting you quite badly, in addition to your choice to enable prefetching (every ZFS FreeBSD system I've used has benefit tremendously from having prefetching disabled, even on systems with 8GB RAM and more). You do not need to specify vm.kmem_size_max, so please remove that. Keeping vm.kmem_size is fine. Also get rid of your vdev tunings, I'm not sure why you have those. My relevant /boot/loader.conf tunings for 8.2-RELEASE (note to readers: the version of FreeBSD you're running, and build date, matters greatly here so do not just blindly apply these without thinking first): # We use Samba built with AIO support; we need this module! aio_load="yes" # Increase vm.kmem_size to allow for ZFS ARC to utilise more memory. vm.kmem_size="8192M" vfs.zfs.arc_max="6144M" # Disable ZFS prefetching # http://southbrain.com/south/2008/04/the-nightmare-comes-slowly-zfs.html # Increases overall speed of ZFS, but when disk flushing/writes occur, # system is less responsive (due to extreme disk I/O). # NOTE: Systems with 8GB of RAM or more have prefetch enabled by # default. vfs.zfs.prefetch_disable="1" # Decrease ZFS txg timeout value from 30 (default) to 5 seconds. This # should increase throughput and decrease the "bursty" stalls that # happen during immense I/O with ZFS. # http://lists.freebsd.org/pipermail/freebsd-fs/2009-December/007343.html # http://lists.freebsd.org/pipermail/freebsd-fs/2009-December/007355.html vfs.zfs.txg.timeout="5" sysctl tunings ======================Please note that the below kern.maxvnodes tuning is based on my system usage, and yours may vary, so you can remove or comment out this option if you wish. The same goes for vfs.ufs.dirhash_maxmem. As for vfs.zfs.txg.write_limit_override, I strongly suggest you keep this commented out for starters; it effectively "rate limits" ZFS I/O, and this smooths out overall performance (otherwise I was seeing what appeared to be incredible network transfer speed, then the system would churn hard for quite some time on physical I/O, then fast network speed, physical I/O, etc... very "bursty", which I didn't want). # Increase send/receive buffer maximums from 256KB to 16MB. # FreeBSD 7.x and later will auto-tune the size, but only up to the max. net.inet.tcp.sendbuf_max=16777216 net.inet.tcp.recvbuf_max=16777216 # Double send/receive TCP datagram memory allocation. This defines the # amount of memory taken up by default *per socket*. net.inet.tcp.sendspace=65536 net.inet.tcp.recvspace=131072 # dirhash_maxmem defaults to 2097152 (2048KB). dirhash_mem has reached # this limit a few times, so we should increase dirhash_maxmem to # something like 16MB (16384*1024). vfs.ufs.dirhash_maxmem=16777216 # # ZFS tuning parameters # NOTE: Be sure to see /boot/loader.conf for additional tunings # # Increase number of vnodes; we've seen vfs.numvnodes reach 115,000 # at times. Default max is a little over 200,000. Playing it safe... kern.maxvnodes=250000 # Set TXG write limit to a lower threshold. This helps "level out" # the throughput rate (see "zpool iostat"). A value of 256MB works well # for systems with 4GB of RAM, while 1GB works well for us w/ 8GB on # disks which have 64MB cache. vfs.zfs.txg.write_limit_override=1073741824 Good luck. -- | Jeremy Chadwick jdc@parodius.com | | Parodius Networking http://www.parodius.com/ | | UNIX Systems Administrator Mountain View, CA, USA | | Making life hard for others since 1977. PGP 4BD6C0CB |
Christer Solskogen
2011-Feb-24 18:46 UTC
ZFS - abysmal performance with samba since upgrade to 8.2-RELEASE
On Thu, Feb 24, 2011 at 8:55 AM, Jeremy Chadwick <freebsd@jdc.parodius.com> wrote:> > ?# Set TXG write limit to a lower threshold. ?This helps "level out" > ?# the throughput rate (see "zpool iostat"). ?A value of 256MB works well > ?# for systems with 4GB of RAM, while 1GB works well for us w/ 8GB on > ?# disks which have 64MB cache. > ?vfs.zfs.txg.write_limit_override=1073741824 > >Sorry if you have said this before, but could you elaborate a bit about this number? For instance, how much does the cache on the disk has to say. In my case: 3x1.5TB raidz with WD15EADS-00R6B0 which has 32MB cache and 12GB memory. What would you recommend and why. -- chs,
Sergey Gavrilov
2011-Mar-18 10:25 UTC
ZFS - abysmal performance with samba since upgrade to 8.2-RELEASE
vfs.zfs.prefetch_disable="1" causes much slow read on sequence data in
my
case.
with vfs.zfs.prefetch_disable="1":
dd if=/pool1/test/idisk1 of=/dev/null bs=1m count=500
524288000 bytes transferred in 18.347177 secs (28575949 bytes/sec)
with vfs.zfs.prefetch_disable="0":
dd if=/pool1/test/idisk1 of=/dev/null bs=1m count=500
524288000 bytes transferred in 3.331806 secs (157358504 bytes/sec)
after few seconds:
dd if=/pool1/test/idisk1 of=/dev/null bs=1m count=500
524288000 bytes transferred in 0.107767 secs (4865009592 bytes/sec)
My system:
FreeBSD 8.2-RELEASE
hw.machine: amd64
hw.model: Intel(R) Xeon(R) CPU E5520 @ 2.27GHz
hw.ncpu: 8
hw.physmem: 17155719168
3ware 9690sa controller
Disks: ST32000445SS
pool: pool1
NAME STATE READ WRITE CKSUM
pool1 ONLINE 0 0 0
raidz2 ONLINE 0 0 0
da12 ONLINE 0 0 0
da11 ONLINE 0 0 0
da10 ONLINE 0 0 0
da9 ONLINE 0 0 0
da16 ONLINE 0 0 0
da15 ONLINE 0 0 0
da14 ONLINE 0 0 0
da13 ONLINE 0 0 0
--
Best regards,
Sergey Gavrilov
Sergey Gavrilov
2011-Mar-21 09:47 UTC
ZFS - abysmal performance with samba since upgrade to 8.2-RELEASE
Yes, I know. But I have a lot of memory and I want it to be used if it improves performance. I tried to test whole system not just fs. And I surprised in why vfs.zfs.prefetch_disable="1" affects so much. 2011/3/18 Paul Mather <paul@gromit.dlib.vt.edu>> > On Mar 18, 2011, at 5:53 AM, Sergey Gavrilov wrote: > > > vfs.zfs.prefetch_disable="1" causes much slow read on sequence data in my > > case. > > > > with vfs.zfs.prefetch_disable="1": > > > > dd if=/pool1/test/idisk1 of=/dev/null bs=1m count=500 > > 524288000 bytes transferred in 18.347177 secs (28575949 bytes/sec) > > > > with vfs.zfs.prefetch_disable="0": > > > > dd if=/pool1/test/idisk1 of=/dev/null bs=1m count=500 > > 524288000 bytes transferred in 3.331806 secs (157358504 bytes/sec) > > > > after few seconds: > > dd if=/pool1/test/idisk1 of=/dev/null bs=1m count=500 > > 524288000 bytes transferred in 0.107767 secs (4865009592 bytes/sec) > > The last dd performance figures are undoubtedly reflecting a read from data > cached in RAM. You should do your test on a file that is bigger than your > RAM size, or at least bigger than the amount of RAM you are dedicating to > ARC. > > Other than that, ZFS prefetching *is* supposed to speed up sequential > accesses, so it's no surprise you notice a speedup in such cases. :-) > > Cheers, > > Paul. > >-- Best regards, Sergey Gavrilov