[Stgt-devel] Tuning iSER for performance

[Pete Wyckoff]

erezz at voltaire.com wrote on Mon, 03 Mar 2008 23:00 +0200:
> > > Another thing that looks strange - from looking at the logs, I don't see
> > > that work is done in parallel. A command is received, iSCSI processing
> > > is done and it is sent to bs processing (which seems to be idle until
> > > that point in time). Now, nothing happens until bs processing is
> > > complete. I will try to look more into the code and try to understand
> > > it, but can someone try to elaborate on this?
> > 
> > sgp_dd issues only one SCSI command at a time?
>  
> We run sgp_dd & disktest with more than a single thread, so I expect to see more than a single cmd in parallel. Note that if I run the same cmd directly on the device (on the target machine), I see better numbers. It means that the benchmark is ok.

Even just a single dd will kick off lots of readahead requests.  The
queue depth is 128 for iser initiator, I think.  If you're not
seeing more than one command at a time in the target, that would be
a problem.

		-- Pete