[openib-general] RE: [PATCH 0/2] opensm: low-level QoS implementation

[Eitan Zahavi]

Hi Sasha, 

I have cut and paste only the sections I would like to further comment
on.

Eitan

> > [EZ] Please note that algorithm to validate the applicability of the
> > above on the
> >          particular fabric is still required as not all devices
support
> > the 16 VLs
> 
> VL numbers are translated according to port's capabilities and
> configured OperVLs (the numbers are MODed).
[EZ] I am not following what you mean here. Can you elaborate?
> 
> > and not all
> >          devices must support VLArb of 8 entries. In such cases we
> > should at least provide
> >           an error describing why the provided setting is
un-realizable.
> 
> In the case of "short" VLArb table the template will be truncated
> (silently) to meet port's capabilities. This is not a error, right?
[EZ] If you do not have an entry for a VL in the VLArb (both high and
low) tables it means this VL will never be scheduled for transmission.
So anybody using this VL will be "blocked". An algorithm to do SL2VL in
such cases can be used to avoid these problems.
> 
> > [EZ] I do not see how the above could be used. Instead I do see
groups
> > of nodes as being
> >          assigned different QoS levels. As we defined "groups of
nodes"
> > in the partition
> >          policy I would propose using the partitions as the means to
> > define node groups.
> > [EZ] So I propose to keep the "trivial" implementation without this
> > level of control.
> 
> At least it does not hurt, and somebody tell me that this will be
> useful. So I would prefer to start with this feature.

[EZ] OK. But in the future the policy will override these default
parameters.
> 
> >          Instead I would prefer having QoS Policy file defined such
that