Re: [PATCH 5/5] IB/srp: Optimize completion queue polling

[Bart Van Assche <bvanassche-HInyCGIudOg@public.gmane.org>]

On 07/08/14 09:55, Or Gerlitz wrote:
> On Fri, Jul 4, 2014 at 12:59 PM, Bart Van Assche <bvanassche-HInyCGIudOg@public.gmane.org> wrote:
>>
>> [...] the blk-iopoll framework defers work to softirq
>> context. This means that a context switch from interrupt to softirq
>> context has to occur before SRP completion processing can start. Recent
>> measurements on current hardware have shown that such a context switch
>> takes about 0.5 microseconds. Since I prefer to keep the latency of SRP
>> I/O as low as possible I haven't looked further into using blk-iopoll for the SRP initiator driver.
> 
> In the same manner that SoftirqD is the right place to process
> received packets in the networking space, it should be
> the case for IO completions in the storage space too.  Note that
> typical packet latency in latest/fastest 10g NICs go below 5us
> when using NAPI so if we (== NIC drivers) are happy with the CS
> overhead there, we (== SCSI LLD storage drivers whose
> latency is in the area of 15-20us) should be happy here too.
> Specifically, the CS cost is amortized across multiple IOs.

Hello Or,

We might each be referring to different concepts here. What you are
referring to is average response time for QD > 1. What I was referring
to is average response time for QD = 1. The measurements I ran with fio
have shown that the response time for QD = 1 is lower when processing
completions in interrupt context compared to processing interrupts in
softirq context.

Bart.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html