Re: slow 'check'

[Bill Davidsen <davidsen@tmr.com>]

Justin Piszcz wrote:
>
>
> On Sat, 10 Feb 2007, Eyal Lebedinsky wrote:
>
>> Justin Piszcz wrote:
>>>
>>>
>>> On Sat, 10 Feb 2007, Eyal Lebedinsky wrote:
>>>
>>>> I have a six-disk RAID5 over sata. First two disks are on the mobo and
>>>> last four
>>>> are on a Promise SATA-II-150-TX4. The sixth disk was added recently
>>>> and I decided
>>>> to run a 'check' periodically, and started one manually to see how
>>>> long it should
>>>> take. Vanilla 2.6.20.
>>>>
>>>> A 'dd' test shows:
>>>>
>>>> # dd if=/dev/md0 of=/dev/null bs=1024k count=10240
>>>> 10240+0 records in
>>>> 10240+0 records out
>>>> 10737418240 bytes transferred in 84.449870 seconds (127145468 
>>>> bytes/sec)
>>>>
>>>> This is good for this setup. A check shows:
>>>>
>>>> $ cat /proc/mdstat
>>>> Personalities : [raid6] [raid5] [raid4]
>>>> md0 : active raid5 sda1[0] sdf1[5] sde1[4] sdd1[3] sdc1[2] sdb1[1]
>>>>      1562842880 blocks level 5, 256k chunk, algorithm 2 [6/6] [UUUUUU]
>>>>      [>....................]  check =  0.8% (2518144/312568576)
>>>> finish=2298.3min speed=2246K/sec
>>>>
>>>> unused devices: <none>
>>>>
>>>> which is an order of magnitude slower (the speed is per-disk, call it
>>>> 13MB/s
>>>> for the six). There is no activity on the RAID. Is this expected? I
>>>> assume
>>>> that the simple dd does the same amount of work (don't we check 
>>>> parity on
>>>> read?).
>>>>
>>>> I have these tweaked at bootup:
>>>>     echo 4096 >/sys/block/md0/md/stripe_cache_size
>>>>     blockdev --setra 32768 /dev/md0
>>>>
>>>> Changing the above parameters seems to not have a significant effect.
>>>>
>>>> The check logs the following:
>>>>
>>>> md: data-check of RAID array md0
>>>> md: minimum _guaranteed_  speed: 1000 KB/sec/disk.
>>>> md: using maximum available idle IO bandwidth (but not more than
>>>> 200000 KB/sec) for data-check.
>>>> md: using 128k window, over a total of 312568576 blocks.
>>>>
>>>> Does it need a larger window (whatever a window is)? If so, can it
>>>> be set dynamically?
>>>>
>>>> TIA
>>>>
>>>> -- 
>>>> Eyal Lebedinsky (eyal@eyal.emu.id.au) <http://samba.org/eyal/>
>>>>     attach .zip as .dat
>>>
>>> As you add disks onto the PCI bus it will get slower.  For 6 disks you
>>> should get faster than 2MB/s however..
>>>
>>> You can try increasing the min speed of the raid rebuild.
>>
>> Interesting - this does help. I wonder why it used much more i/o by
>> default before. It still uses only ~16% CPU.
>>
>> # echo 20000 >/sys/block/md0/md/sync_speed_min
>> # echo check >/sys/block/md0/md/sync_action
>> ... wait about 10s for the process to settle...
>> # cat /proc/mdstat
>> Personalities : [raid6] [raid5] [raid4]
>> md0 : active raid5 sda1[0] sdf1[5] sde1[4] sdd1[3] sdc1[2] sdb1[1]
>>      1562842880 blocks level 5, 256k chunk, algorithm 2 [6/6] [UUUUUU]
>>      [>....................]  check =  0.1% (364928/312568576) 
>> finish=256.6min speed=20273K/sec
>> # echo idle >/sys/block/md0/md/sync_action
>>
>> Raising it further only manages about 21MB/s (the _max is set to 
>> 200MB/s)
>> as expected; this is what the TX4 delivers with four disks. I need a 
>> better
>> controller (or is the linux driver slow?).
>>
>>> Justin.
>
> You are maxing out the PCI Bus, remember each bit/parity/verify 
> operation has to go to each disk.  If you get an entirely PCI-e system 
> you will see rates 50-100-150-200MB/s easily.  I used to have 10 x 
> 400GB drives on a PCI bus, after 2 or 3 drives, you max out the PCI 
> bus, this is why you need PCI-e, each slot has its own lane of bandwidth.

>
> 21MB/s is about right for 5-6 disks, when you go to 10 it drops to 
> about 5-8MB/s on a PCI system. 
Wait, let's say that we have three drives and 1m chunk size. So we read 
1M here, 1M there, and 1M somewhere else, and get 2M data and 1M parity 
which we check. With five we would read 4M data and 1M parity, but have 
4M checked. The end case is that for each stripe we read N*chunk bytes 
and verify (N-1)*chunk. In fact the data is (N-1)/N of the stripe, and 
the percentage gets higher (not lower) as you add drives. I see no 
reason why more drives would be slower, a higher percentage of the bytes 
read are data.

That doesn't mean that you can't run out of Bus bandwidth, but number of 
drives is not obviously the issue.

-- 
bill davidsen <davidsen@tmr.com>
  CTO TMR Associates, Inc
  Doing interesting things with small computers since 1979