[ofa-general] Re: [PATCH 1/4] ib/ipoib: specify TClassand FlowLabelwith PR queries for QoS support

[Michael S. Tsirkin]

> Quoting Sean Hefty <mshefty at ichips.intel.com>:
> Subject: Re: [ofa-general] Re: [PATCH 1/4] ib/ipoib: specify?TClassand?FlowLabelwith PR queries for QoS support
> 
> >AFAIK tavor only works better with 1K MTU with connected transport types.
> >With UD, the overhead of extra WRs seems to outweight the speedup
> >we get from hardware.
> 
> Doesn't IPoIB use the same path record for both UD and RC, if RC is 
> available?

Good point, yes it does.

> I guess the question still remains, does IPoIB need to (or should it) 
> specify the MTU, or should it be set by the SA?  It sounds like IPoIB 
> can handle the case where the path MTU is less than the device MTU.

Here's what happens if we activate tavor workaround in SM:
ipoib datagram mode actually *ignores* the MTU from path query
and uses the MTU from broadcast group. ipoib CM uses MTU
from path query.

As a result, datagram mode uses 2K MTU while connected mode uses 1K MTU.

> It doesn't specify it today, and the QoS patches seem like a somewhat 
> separate issue.

I agree.

-- 
MST