RFC: GEOM MULTIPATH rewrite

[Alexander Motin]

On 01/20/12 14:13, Nikolay Denev wrote:
> On Jan 20, 2012, at 1:30 PM, Alexander Motin wrote:
>> On 01/20/12 13:08, Nikolay Denev wrote:
>>> On 20.01.2012, at 12:51, Alexander Motin<mav at freebsd.org>   wrote:
>>>
>>>> On 01/20/12 10:09, Nikolay Denev wrote:
>>>>> Another thing I've observed is that active/active probably only makes sense if you are accessing single LUN.
>>>>> In my tests where I have 24 LUNS that form 4 vdevs in a single zpool, the highest performance was achieved
>>>>> when I split the active paths among the controllers installed in the server importing the pool. (basically "gmultipath rotate $LUN" in rc.local for half of the paths)
>>>>> Using active/active in this situation resulted in fluctuating performance.
>>>>
>>>> How big was fluctuation? Between speed of one and all paths?
>>>>
>>>> Several active/active devices without knowledge about each other with some probability will send part of requests via the same links, while ZFS itself already does some balancing between vdevs.
>>>
>>> I will test in a bit and post results.
>>>
>>> P.S.: Is there a way to enable/disable active-active on the fly? I'm
>>> currently re-labeling to achieve that.
>>
>> No, there is not now. But for experiments you may achieve the same results by manually marking as failed all paths except one. It is not dangerous, as if that link fail, all other will resurrect automatically.
>
> I had to destroy and relabel anyways, since I was not using active-active currently. Here's what I did (maybe a little too verbose):
>
> And now a very naive benchmark :
>
> :~# dd if=/dev/zero of=/tank/TEST bs=1M count=512
> 512+0 records in
> 512+0 records out
> 536870912 bytes transferred in 7.282780 secs (73717855 bytes/sec)
> :~# dd if=/dev/zero of=/tank/TEST bs=1M count=512
> 512+0 records in
> 512+0 records out
> 536870912 bytes transferred in 38.422724 secs (13972745 bytes/sec)
> :~# dd if=/dev/zero of=/tank/TEST bs=1M count=512
> 512+0 records in
> 512+0 records out
> 536870912 bytes transferred in 10.810989 secs (49659740 bytes/sec)
>
> Now deactivate the alternative paths :
> And the benchmark again:
>
> :~# dd if=/dev/zero of=/tank/TEST bs=1M count=512
> 512+0 records in
> 512+0 records out
> 536870912 bytes transferred in 1.083226 secs (495622270 bytes/sec)
> :~# dd if=/dev/zero of=/tank/TEST bs=1M count=512
> 512+0 records in
> 512+0 records out
> 536870912 bytes transferred in 1.409975 secs (380766249 bytes/sec)
> :~# dd if=/dev/zero of=/tank/TEST bs=1M count=512
> 512+0 records in
> 512+0 records out
> 536870912 bytes transferred in 1.136110 secs (472551848 bytes/sec)
>
> P.S.: The server is running 8.2-STABLE, dual port isp(4) card, and is directly connected to a 4Gbps Xyratex dual-controller (active-active) storage array.
> All the 24 SAS drives are setup as single disk RAID0 LUNs.

This difference is too huge to explain it with ineffective paths 
utilization. Can't this storage have some per-LUN port/controller 
affinity that may penalize concurrent access to the same LUN from 
different paths? Can't it be active/active on port level, but 
active/passive for each specific LUN? If there are really two 
controllers inside, they may need to synchronize their caches or bounce 
requests, that may be expensive.

-- 
Alexander Motin