[ofa-general] SRP/mlx4 interrupts throttling performance

[Cameron Harr]

Vladislav Bolkhovitin wrote:
> Cameron Harr, on 01/09/2009 02:10 AM wrote:
>> Also, I think my original question got lost in the main thread, which 
>> was to find out why I could get double the performance exporting two 
>> drives over SCST/SRP as opposed to raiding those two drives and 
>> exporting them as one object.
>
> Hmm, I can't see where you asked it. I checked your original e-mail 
> and didn't find it.
You're absolutely right. Looking back at the original email, I didn't 
really say the question. I had tried debugging the issue on my own and 
determined interrupts were the problem and made that be my question. I 
never got back to asking the base question.

>
> Anyway, can you repeat your question? Unfortunately, from the above I 
> can't understand it.
>
Back when I started the thread, I had the problem where performance 
seemed to follow the number of devices exported instead of the 
performance of the devices themselves. Specifically, I had two drives on 
a local box. If I did a RAID 0 on those two drives and exported them, 
I'd get performance X. However, if I exported both of the drives as 
individual drives, my performance was close to 2X.

Just now, running with recent code, I ran a battery of tests to verify 
with real numbers, and to my pleasant surprise, the issue (which had 
been verified by multiple people at my work) seems to have vanished. I'm 
now getting equal performance out of 1 2-drive RAID SRP target as I am 
out of 2 independent SRP targets. So, congrats - you fixed my problem!