[dm-devel] Performance issue when building thin-pool on top of RAID6 device

[Zdenek Kabelac]

Dne 18.2.2014 11:36, Teng-Feng Yang napsal(a):
> Hi,
>
> Sorry, I forget to provide more information on my system configuration.
> I use Ubuntu 13.10 and update linux kernel to 3.13.3.
>
> The dmsetup table output should be similar in all experiments as follow:
> thin: 0 80000000 thin 252:0 0
> pool: 0 117433344 thin-pool 8:145 9:0 2048 0 1 skip_block_zeroing
>

Ok - now - have you tried other block size except for 1MB ?

I'd suggest to fit the  block size into raid6 multiple chunk size.
So something like reduce chunk_size in raid6 to 128KiB (or maybe even 64KiB) 
per disk and use  640KiB as a block size for thin pool ?

Several variants are probably worth to check here.

Thanks

Zdenek




> Thanks,
> Dennis
>
> 2014-02-18 17:34 GMT+08:00 Zdenek Kabelac <zkabelac at redhat.com>:
>> Dne 18.2.2014 10:13, Teng-Feng Yang napsal(a):
>>
>>> Dear all,
>>>
>>> I have been working on tuning the dm-thin performance on my storage
>>> server.
>>> I surprisingly find out that the write performance of newly created
>>> thin volume formatted as EXT4 has degraded significantly, when
>>> building thin-pool on top of RAID6 device.
>>> Thin pools in all experiment use 1MB block size, and I format all
>>> target volumes as EXT4 and mount them on /mnt.
>>> The following is the performance statistics I gather under different
>>> circumstances.
>>>
>>> Thin volume on top of a thin pool created by a Plextor M5P 128GB SSD
>>> as metadata device
>>> and a RAID6 block device composed by 7 disks (chunk size = 512KB) as a
>>> data device
>>> dennis at desktop:~$ sudo dd if=/dev/zero of=/mnt/zero.img bs=1M count=25000
>>> 25000+0 records in
>>> 25000+0 records out
>>> 26214400000 bytes (26 GB) copied, 245.808 s, 107 MB/s
>>>
>>> Directly make filesystem on RAID6 block device composed by 7 disks
>>> dennis at desktop:~$ sudo dd if=/dev/zero of=/mnt/zero.img bs=1M count=25000
>>> 25000+0 records in
>>> 25000+0 records out
>>> 26214400000 bytes (26 GB) copied, 129.543 s, 202 MB/s
>>>
>>> Thin volume on top of a thin pool created by a Plextor M5P 128GB SSD
>>> as metadata device
>>> and a RAID0 block device composed by 7 disks as a data device
>>> dennis at desktop:~$ sudo dd if=/dev/zero of=/mnt/zero.img bs=1M count=25000
>>> 25000+0 records in
>>> 25000+0 records out
>>> 26214400000 bytes (26 GB) copied, 46.1227 s, 568 MB/s
>>>
>>> Directly make filesystem on RAID6 block device composed by 7 disks
>>> dennis at desktop:~$ sudo dd if=/dev/zero of=/mnt/zero.img bs=1M count=25000
>>> 25000+0 records in
>>> 25000+0 records out
>>> 26214400000 bytes (26 GB) copied, 48.1104 s, 545 MB/s
>>>
>>> It is clear that the write performance has degraded significantly when
>>> we use RAID6 as thin-pool,
>>> but there is no similar behavior when using RAID0 as thin-pool instead.
>>> I dig a little deeper on this issue and find that if we can perfectly
>>> fit a pool block into a RAID6 stripe,
>>> there would be only 20~30% performance loss comparing to RAW RAID6.
>>> However, this poses a restriction on the disks number we can use to
>>> build a RAID6 for pool, and limit the maximum pool size as well.
>>> Is there any other way I can do to improve the write performance when
>>> using RAID6 as pool?
>>
>>
>>
>> Since you have not provided any  'dmsetup table' output - it's hard to guess
>> how your thin-pool target is created, but my guess is - you are using
>> 'zeroing' for provisioned blocks.
>>
>> i.e. here is some sample for thin-pool target line with disabled zeroing:
>>
>> vg-pool-tpool: 0 40960 thin-pool 253:1 253:2 128 0 1 skip_block_zeroing
>>
>>
>> (and example of lvm2 command line:  'lvcreate -T -L100G -Zn -c1M vg/pool'
>>
>> Zdenek
>>
>
> --
> dm-devel mailing list
> dm-devel at redhat.com
> https://www.redhat.com/mailman/listinfo/dm-devel
>