[sheepdog] Redundancy policy via iSCSI

hujianyang hujianyang at huawei.com
Sat Feb 7 03:03:18 CET 2015


On 2015/2/6 16:41, Hitoshi Mitake wrote:
> At Wed, 04 Feb 2015 11:24:21 +0900,
> Hitoshi Mitake wrote:
>>
>> At Tue, 3 Feb 2015 17:17:42 +0800,
>> hujianyang wrote:
>>>
>>> Hi Saeki,
>>>
>>> On 2015/2/3 16:53, Saeki Masaki wrote:
>>>> Hi Hu,
>>>>
>>>> Since Sheepdog has a mechanism that does not place objects in the same zone_id.
>>>> Can you try to change ZONE id in each node.
>>>>>>>    Id   Host:Port         V-Nodes       Zone
>>>>>>>     0   130.1.0.147:7000        128          0
>>>>>>>     1   130.1.0.148:7000        128          0
>>>>>>>     2   130.1.0.149:7000        128          0
>>>>
>>>> Best Regards, Saeki.
>>>>
>>>
>>> Good suggestions~!
>>>
>>> Seems OK now. But write performance is too slow in my environment.
>>
>> 1.1MB/s seems to be too slow, how about changing input file from
>> /dev/random to /dev/zero? And I'd like to know perofrmance of default
>> backing store of tgt (use file as iSCSI target) on your environment.
>>
>> Thanks,
>> Hitoshi
> 
> BTW, I have pending patchset for parallelizing iSCSI PDU send/recv of
> tgtd:
> https://github.com/mitake/tgt/commits/iscsi-pdu-rxtx-mt
> 
> You can activate the feature with new option -T:
> $ tgtd -T 16
> 
> It is still half-baked, but in some cases it can improve performance
> of iSCSI + sheepdog.
> 
> Thanks,
> Hitoshi
> 

Hi Hitoshi,

Sorry for reply late. You know, there always many stuffs need to
been done before Spring Festival.

Actually my current environment is just for testing the features
of sheepdog. Performance is not a urgent issue. Thanks for your
kindness.

I have tested sheepdog with fio on my testing environment:

[global]
runtime=300
direct=1
iodepth=1
bs=256K
size=100G
numjobs=1
time_based

KB/s			read	randread	write	randwrite
local			179957	36262		179933	66752
iSCSI	redundancy(3x)	51553	51303		17826	15984
	redundancy(4:2)	43166	42775		20370	14234
SBD	redundancy(3x)	51112	51106		17515	20311


I'm not quite sure why randwrite is better than randread via local
access.

Thanks,
Hu




More information about the sheepdog mailing list