[Stgt-devel] Tuning iSER for performance

[Erez Zilber]

>> It seems that bs work takes a lot of time. For example: if I run
>> disktest on the same device directly from the target machine (i.e. tgt
>> is not involved) in the following way:
>>
>> disktest -PT -T100 -h1 -K64 -B256k -ID /dev/sdc -r
>>
>> I get ~13000 iops -> each IO takes ~77 us (compared to 130 us only for
>> bs work with tgt). I'm not familiar with the bs code in tgt, so I'm not
>> sure what should be done here.
>>     
>
> The problem is that Linux lacks a nice event notification
> mechanism. The main thread uses epoll to wait on events. I/O threads
> (pthread) use pipe to notify the main thread of I/O completion. It's
> not effective at all.
>
>   

If I understand correctly, it sounds like a major problem in stgt (that
also affects other protocols - not only iSCSI). When a SCSI cmd is done,
it takes time until the iSCSI (or any other protocol) layer is notified,
correct? Is there a plan or ideas on how to solve that?

BTW - I don't see any connection between the problem that you describe
and the problem that SCSI executes a single command at a time (host_busy
<= 1). As I said, I don't think that the problem is sgp_dd (that runs
with thr=8).

Erez