Re: [Scst-devel] nv_cache performance

[Alexey <alu...@al...>]

Hello, Bart.

On 2015-06-10 03:35, Bart Van Assche wrote:
> On 06/09/15 08:11, Alexey wrote:
>> I have storage node with linux 3.10.61, scst 3.0.1+svn5410.
>>      and compute node with linux 3.19.1,  scst 3.0.1+svn5410, srptools
>> 1.0.2.
> 
> Hello Alexey,
> 
> Have you already had the chance to read the srpt/README document ? In
> that document there is a section about optimizing initiator systems
> for performance and another section about optimizing target systems
> for performance.

I recompile SCST from svn revision 6135.
I reread README and played again with some kernel parameters like 
readahead, max_sectors_kb and nr_requests.
After new tests I was able to increase randread performance from 
nv_cache up to 195kiops... But it is not enough for me.
Initiator connected to the target with 40GB/s Infiniband cable.

During my observations, I have never seen a spike above 200k iops. And I 
find it very strange.
All tests were made with 4k block size.

Regards,
Alexey Mochkin.