Skip to contentSkip to conversation

3ware disk latency?

DISCUSSION Jan Kasprzak 7 participants 12 messages Jul 10, 2006 – Jul 27, 2006

lore
Share & tools
lore
#0 lore raw

Hi all,

I have upgraded my machine from 3ware 7508 with 8x 250GB ATA drives to 3ware 9550SX-8LP with 8x 500GB SATA-II drives, and I have found that while the overall throughput of the disks is higher than before, the latency - measured as the number of messages per time unit Qmail can deliver - is far worse than before[1]. I have tried to use deadline and cfq schedulers, but it made no visible speedup.

I have found the following two years old mail by Jens Axboe:
http://www.uwsg.iu.edu/hypermail/linux/kernel/0409.0/1330.html
- the problem at that time was that the device had its internal command queue deeper than the nr_requests of the block device layer, so the I/O scheduler couldn't do anything. I was surprised that the situation is still the same: I have

/sys/block/sd[a-h]/queue/nr_requests == 128, and /sys/devices/pci0000:00/0000:00:0b.0/0000:01:03.0/host0/target*/*/queue_depth == 254

I have verified that even on older 3ware drivers (7508) the situation is the same. On the newer hardware it is probably more visible, because the controller actually has bigger on-board cache (128MB vs. 32MB, I think).

I have tried to lower the /sys/devices/.../queue_depth to 4
and disable the NCQ, and the latency got a bit better. But it is still nowhere near to where it was on the older HW.

Does anybody experience the similar latency problems with
3ware 95xx controllers? Thanks,

-Yenya

[1] the old configuration peaked at ~2k-4k messages per 5 minutes, with 1k-2k messages/5min being pretty normal, while the new one has maximum throughput of 1k messages per 5 minutes, but the normal speed is much lower - some low hundreds messages per 5 minutes. The new system runs about the same load as the previous one, and the layout of disks is also the same (just the newer drives are bigger, of course).

Signature · 3 lines
| Jan "Yenya" Kasprzak  <kas at {fi.muni.cz - work | yenya.net - private}> |
| GPG: ID 1024/D3498839      Fingerprint 0D99A7FB206605D7 8B35FCDE05B18A5E |
| http://www.fi.muni.cz/~kas/    Journal: http://www.fi.muni.cz/~kas/blog/ |
Quoted 2 lines

I will never go to meetings again because I think face to face meetings <
are the biggest waste of time you can ever have. --Linus Torvalds <

#1 lore raw
Quoted 2 lines · Jan Kasprzak

Does anybody experience the similar latency problems with
3ware 95xx controllers? Thanks,

i think the 95xx is really bad at sharing its cache fairly amongst multiple devices... i have a 9550sx-8lp with 3 exported units: a raid1, a raid10 and a jbod. i was zeroing the jbod with dd and it absolutely destroyed the latency of the other two units.

to make things sane i had to add oflag=direct to dd to avoid dirtying all of memory... and i had to tell the 3ware to disable write caching on the jbod device.

the result is a fairly slow throughput on the dd (20MB/s) ... but at least the latency of the raid1 and raid10 aren't being affected any more.

i suspect that telling the 9550sx to not use nvram limits the amount of its tiny 128MiB of RAM which can be consumed by that unit.

having to use oflag=direct indicates linux is at least partially at fault here... maybe there's some tunable i need to learn about... the system has 8GiB of ram, maybe the default dirty ratio of 40% is too high. but i suspect the problem is more that with dd dirtying memory immediately it causes the system to be stuck at the 40% dirty watermark which causes excessive write traffic even to the raid1/raid10 units because they too are contributing to the 40% dirty (temporary files and whatnot would hit disk when they wouldn't in normal operation).

i think since i'm using 3ware hw raid on this setup with only two units normally active it isn't as painful as what you're seeing with what looks like 8 JBOD units.

yeah i'm a bit disappointed... i neglected to benchmark this sort of concurrency thing when i was experimenting with the 9550sx before turning it live. i'm curious if areca controllers handle this better.

-dean

#2 lore raw

Ar Mer, 2006-07-26 am 09:52 -0700, ysgrifennodd dean gaudet:

Quoted 9 lines · Jan Kasprzak
Quoted 3 lines

Does anybody experience the similar latency problems with
3ware 95xx controllers? Thanks,

i think the 95xx is really bad at sharing its cache fairly amongst multiple devices... i have a 9550sx-8lp with 3 exported units: a raid1, a raid10 and a jbod. i was zeroing the jbod with dd and it absolutely destroyed the latency of the other two units.

Even on the older 3ware I found it neccessary to bump up the queue sizes for the 3ware. There's a thread from about 6 months ago (maybe a bit more) with Jeff Merkey included where it made a huge difference to his performance playing with the parameters. Might be worth a scan through the archive.

Alan

#3 lore raw
Quoted 11 lines · Alan Cox

Ar Mer, 2006-07-26 am 09:52 -0700, ysgrifennodd dean gaudet:

Quoted 9 lines · Jan Kasprzak
Quoted 3 lines

Does anybody experience the similar latency problems with
3ware 95xx controllers? Thanks,

i think the 95xx is really bad at sharing its cache fairly amongst multiple devices... i have a 9550sx-8lp with 3 exported units: a raid1, a raid10 and a jbod. i was zeroing the jbod with dd and it absolutely destroyed the latency of the other two units.

Quoted 5 lines

Even on the older 3ware I found it neccessary to bump up the queue sizes for the 3ware. There's a thread from about 6 months ago (maybe a bit more) with Jeff Merkey included where it made a huge difference to his performance playing with the parameters. Might be worth a scan through the archive.

yeah i've done some experiments with a multithreaded random O_DIRECT i/o generator where i showed the 9550sx peaking at around 192 threads... where "peak" means the latency was still reasonable (measured by iostat). obviously when going past 128 threads i found it necessary to increase nr_requests to see an improvement in io/s. there was some magic number past 256 where i found the 3ware must have choked its own intnernal queue and latency became awful (istr it being 384ish but it probably depends on the io workload).

unfortunately when i did the experiment i neglected to perform simultaneous tests on more than one 3ware unit on the same controller. i got great results from a raid1 or from a raid10 (even a raid5)... but never i only tested one unit at a time.

my production config has a raid1 and a raid10 unit on the same controller -- with 2 spare ports. i popped a disk into one of the spare ports, told 3ware it was JBOD and then proceeded to "dd if=/dev/zero of=/dev/sdX bs=1M"... and wham my system load avg shot up to 80+ (from ~4), all in D state.

i tried a bunch of things... lowering and raising nr_requests on all the devices, changing between cfq, deadline, etc. the only solution i found to be effective in retaining acceptable latencies on the raid1/raid10 was to use "oflag=direct" for dd, and tell the 3ware to not use nvram for the jbod disk.

my suspicion is the 3ware lacks any sort of "fairness" in its sharing of buffer space between multiple units on the same controller. and by disabling the write caching it limits the amount of controller memory that the busy disk can consume.

-dean

#4 lore raw
Quoted 6 lines · dean gaudet

unfortunately when i did the experiment i neglected to perform simultaneous tests on more than one 3ware unit on the same controller. i got great results from a raid1 or from a raid10 (even a raid5)... but never i only tested one unit at a time.

Did you try setting /sys/class/scsi_host/hostX/cmd_per_lun to 256 / num_units ?

-Adam

#5 lore raw
Quoted 12 lines · adam radford
Quoted 6 lines · dean gaudet

unfortunately when i did the experiment i neglected to perform simultaneous tests on more than one 3ware unit on the same controller. i got great results from a raid1 or from a raid10 (even a raid5)... but never i only tested one unit at a time.

Did you try setting /sys/class/scsi_host/hostX/cmd_per_lun to 256 / num_units ?

(quoted diff) +0 -2
-Adam
-
Quoted 5 lines

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Speaking of this does anyone having any tuning guides for running 10 disk raid 5's on this controller? And as far as ext3 goes, what are the best optimizations?

#6 lore raw
In reply to adam radford
Quoted 13 lines · adam radford
Quoted 6 lines · dean gaudet

unfortunately when i did the experiment i neglected to perform simultaneous tests on more than one 3ware unit on the same controller. i got great results from a raid1 or from a raid10 (even a raid5)... but never i only tested one unit at a time.

Did you try setting /sys/class/scsi_host/hostX/cmd_per_lun to 256 / num_units ?

(quoted diff) +0 -2
-Adam
-
Quoted 5 lines

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Adam,

Isn't it set to this by default or is the default 128?

Jeff

#7 lore raw
Quoted 3 lines · Jeff V. Merkey

Isn't it set to this by default or is the default 128?

Default is 254.

-Adam

#8 lore raw
In reply to adam radford
Quoted 10 lines · adam radford
Quoted 6 lines · dean gaudet

unfortunately when i did the experiment i neglected to perform simultaneous tests on more than one 3ware unit on the same controller. i got great results from a raid1 or from a raid10 (even a raid5)... but never i only tested one unit at a time.

Did you try setting /sys/class/scsi_host/hostX/cmd_per_lun to 256 / num_units ?

hmm doesn't look like i can do this in 2.6.16.27:

# ls -l /sys/class/scsi_host/host0/cmd_per_lun
-r--r--r-- 1 root root 0 Jul 26 19:58 /sys/class/scsi_host/host0/cmd_per_lun
# echo 85 >!$
echo 85 >/sys/class/scsi_host/host0/cmd_per_lun
zsh: permission denied: /sys/class/scsi_host/host0/cmd_per_lun

it'll probably be sometime until i update this box past 2.6.16.x ... but i'll keep it in mind and retry the experiment... thanks.

-dean

#9 lore raw

dean gaudet wrote:
: On Wed, 26 Jul 2006, adam radford wrote:
: >
: > Did you try setting /sys/class/scsi_host/hostX/cmd_per_lun to 256 / num_units : > ?
:
: hmm doesn't look like i can do this in 2.6.16.27:
:
: # ls -l /sys/class/scsi_host/host0/cmd_per_lun
: -r--r--r-- 1 root root 0 Jul 26 19:58 /sys/class/scsi_host/host0/cmd_per_lun : # echo 85 >!$
: echo 85 >/sys/class/scsi_host/host0/cmd_per_lun
: zsh: permission denied: /sys/class/scsi_host/host0/cmd_per_lun
:
: it'll probably be sometime until i update this box past 2.6.16.x ... but : i'll keep it in mind and retry the experiment... thanks.

2.6.17.7 also does not allow cmd_per_lun to be written.

-Yenya

Signature · 3 lines
| Jan "Yenya" Kasprzak  <kas at {fi.muni.cz - work | yenya.net - private}> |
| GPG: ID 1024/D3498839      Fingerprint 0D99A7FB206605D7 8B35FCDE05B18A5E |
| http://www.fi.muni.cz/~kas/    Journal: http://www.fi.muni.cz/~kas/blog/ |
Quoted 2 lines

I will never go to meetings again because I think face to face meetings <
are the biggest waste of time you can ever have. --Linus Torvalds <

#10 lore raw
In reply to dean gaudet

dean gaudet wrote:
: my suspicion is the 3ware lacks any sort of "fairness" in its sharing of : buffer space between multiple units on the same controller. and by : disabling the write caching it limits the amount of controller memory that : the busy disk can consume.

Hmm, do you have a battery backup unit for 9550sx? I don't,
and the 3ware BIOS does not even allow me to enable write caching without it. So I don't think the write caching on the controller side is related to my problem.

I have been able to improve the latency by upgrading the firmware
to the newest release (wow, they even have a firmware updating utility for Linux!).

-Yenya

Signature · 3 lines
| Jan "Yenya" Kasprzak  <kas at {fi.muni.cz - work | yenya.net - private}> |
| GPG: ID 1024/D3498839      Fingerprint 0D99A7FB206605D7 8B35FCDE05B18A5E |
| http://www.fi.muni.cz/~kas/    Journal: http://www.fi.muni.cz/~kas/blog/ |
Quoted 2 lines

I will never go to meetings again because I think face to face meetings <
are the biggest waste of time you can ever have. --Linus Torvalds <

#11 lore raw
Quoted 10 lines · Jan Kasprzak

dean gaudet wrote:
: my suspicion is the 3ware lacks any sort of "fairness" in its sharing of : buffer space between multiple units on the same controller. and by : disabling the write caching it limits the amount of controller memory that : the busy disk can consume.

Hmm, do you have a battery backup unit for 9550sx? I don't,
and the 3ware BIOS does not even allow me to enable write caching without it. So I don't think the write caching on the controller side is related to my problem.

interesting, on the 9500 series cards I have it pops up a 'are you really sure, this is dangerous' warning, but it did allow me to turn on write caching (I've since received the batteries, but I did test it before that)

David Lang

Quoted 7 lines

I have been able to improve the latency by upgrading the firmware
to the newest release (wow, they even have a firmware updating utility for Linux!).

-Yenya