Re: Raid-5 long write wait while reading

[Bill Davidsen <davidsen@tmr.com>]

tj wrote:
> Bill Davidsen wrote:
>> tj wrote:
>>> Thomas Jager wrote:
>>>> Hi list.
>>>>
>>>> I run a file server on MD raid-5.
>>>> If a client reads one big file and at the same time another client 
>>>> tries to write a file, the thread writing just sits in 
>>>> uninterruptible sleep until the reader has finished. Only very 
>>>> small amount of writes get trough while the reader is still working.
>>>> I'm having some trouble pinpointing the problem.
>>>> It's not consistent either sometimes it works as expected both the 
>>>> reader and writer gets some transactions. On huge reads I've seen 
>>>> the writer blocked for 30-40 minutes without any significant writes 
>>>> happening (Maybe a few megabytes, of several gigs waiting). It 
>>>> happens with NFS, SMB and FTP, and local with dd. And seems to be 
>>>> connected to raid-5. This does not happen on block devices without 
>>>> raid-5. I'm also wondering if it can have anything to do with 
>>>> loop-aes? I use loop-aes on top of the md, but then again i have 
>>>> not observed this problem on loop-devices with disk backend. I do 
>>>> know that loop-aes degrades performance but i didn't think it would 
>>>> do something like this?
>>>>
>>>> I've seen this problem in 2.6.16-2.6.21
>>>>
>>>> All disks in the array is connected to a controller with a SiI 3114 
>>>> chip.
>>>
>>> I just noticed something else. A couple of slow readers where 
>>> running on my raid-5 array. Then i started a copy from another local 
>>> disk to the array. Then i got the extremely long wait. I noticed 
>>> something in iostat:
>>>
>>> avg-cpu:  %user   %nice %system %iowait  %steal   %idle
>>>           3.90    0.00   48.05   31.93    0.00   16.12
>>>
>>> Device:            tps    kB_read/s    kB_wrtn/s    kB_read    kB_wrtn
>>> ....
>>> sdg               0.80        25.55         0.00        128          0
>>> sdh             154.89       632.34         0.00       3168          0
>>> sdi               0.20        12.77         0.00         64          0
>>> sdj               0.40        25.55         0.00        128          0
>>> sdk               0.40        25.55         0.00        128          0
>>> sdl               0.80        25.55         0.00        128          0
>>> sdm               0.80        25.55         0.00        128          0
>>> sdn               0.60        23.95         0.00        120          0
>>> md0             199.20       796.81         0.00       3992          0
>>>
>>> All disks are member of the same raid array (md0). One of the disks 
>>> has a ton of transactions compared to the other disks. Read 
>>> operations as far as i can tell. Why? May be connected with my problem? 
>> Two thoughts on that, if you are doing a lot of directory operations, 
>> it's possible that the inodes being used most are all in one chunk.
> Hi thanks for the reply.
>
> It's not directory operations AFAIK. Reading a few files (3 in this 
> case) and writing one.
>>
>> The other possibility is that these a journal writes and reflect 
>> updates to the atime. The way to see if this is in some way  related 
>> is to mount (remount) with noatime: "mount -o remount,noatime 
>> /dev/md0 /wherever" and retest. If this is journal activity you can 
>> do several things to reduce the problem, which I'll go into (a) if it 
>> seems to be the problem, and (b) if someone else doesn't point you to 
>> an existing document or old post on the topic. Oh, you could also try 
>> mounting the filesystem as etc2, assuming that it's ext3 now. I 
>> wouldn't run that way, but it's useful as a diagnostic tool.
> I don't use ext3 i use ReiserFS. ( It seemed like a good idea at the 
> time. ) It's mounted with  -o  noatime.
> I've done some more testing and i seems like it might be connected to 
> mount --bind. If i write to a binded mount i get the slow writes. But 
> if i write directly to the real mount i don't. It might just be a 
> random occurrence, as the problem always has been inconsistent. Thoughts? 

I said I would test, and I did. I don't see a difference with ext3 in 
reads at all. I don't see a difference in bind vs. direct for write, 
either, but all of my space large enough to have room for a few GB write 
had internal bitmaps.

Other info: block size made no consistent difference, changing the 
stripe_cache_size helped but was very non-linear in effect, and direct 
raid over partitions had the same performance as LVM over raid on other 
partitions of the same disks.

Neil: is there a reason (other than ease of coding) why the bitmap isn't 
distributed to minimize seeks? ie. put the bitmap for given stripes at 
the end of those strips rather than the end of the space.

I have added to my tests-to-do list partitioning a disk such that I have 
a small partition and a large, making RAID-10 no bitmap on the small 
(multiple drive, obviously) and then RAID-5 on the large, with the 
bitmap for RAID-5 on the RAID-10 raw array. The only reason I have any 
interest is that I did something like this with JFS, putting the journal 
on a dedicated partition with a different chunk size, and it really 
helped. If this gives any useful information I'll report, but I'm 
building a few analysis tools first, so it will be several weeks.

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