Mindaugas Riauba wrote:
>>>   By the way did you try to test random read performance and latency
>>> of XIV? Looking into architecture (multiple hosts with 7200rpm SATA
>>> drives interconnected via 1Gb Ethernet) it should not perform well in
>>> this area.
>>
>> Yes: When I test with fio (6 parallel tasks, aio, direct I/O, depth
>> 256, 70% reads), the XIV beats the local RAID by a _wide_ margin.
> 
>   Could you try 1 task, 100% random read? And give us results? :)

Yes, but I'm not sure which utility to use to create 100% random reads. 
Perhaps fio in a 100% read mode?


>> I know that the XIV's focus is on parallel I/O, but still: I'm
>> wondering _where_ in the system the bottleneck appears (knowing that I
>> can get fine throughput if I run (e.g.) six dd/cat processes in
>> parallel).
> 
>   XIV is the bottleneck. As I already wrote - SAN storage boxes are
> tuned so that they will not give their full power to 1 I/O process.

But XIV can't see processes on the Linux box. So I reckon it's the 
operating system's duty to spread the I/O over the available paths, as 
long as there aren't other factors (such as CPU bottlenecks) standing in 
the way? This, of course, assumes an active-active(-active)*-situation.


>   But you can look into alignment issues. If I remember right XIV
> splits data into 1MB chunks but I don't know what block size is used at
> lower level.

True, its chunks are 1MB. But what would I align? Setting ext3's block 
size to 1MB seems like a rather weird idea.


> Also you can try to tune read-ahead settings on XIV if that
> is possible.

I don't think that's possible, but I tried it on the OS side (blockdev --
setra ...) without seeing any change.

-- 
Regards,
Troels Arvin <[email protected]>
http://troels.arvin.dk/

_______________________________________________
rhelv5-list mailing list
[email protected]
https://www.redhat.com/mailman/listinfo/rhelv5-list

Reply via email to