[ofa-general] Re: rate assignment for path queries

Hal Rosenstock hal.rosenstock at gmail.com
Sat Nov 15 02:34:38 PST 2008


On Fri, Nov 14, 2008 at 4:39 PM, Hal Rosenstock
<hal.rosenstock at gmail.com> wrote:
> On Thu, Nov 13, 2008 at 8:43 AM, Or Gerlitz <ogerlitz at voltaire.com> wrote:
>> Sasha Khapyorsky wrote:
>>>>
>>>> RED=0x8001, ipoib, sl=1, rate=2, defmember=full : ALL=full;
>>>>
>>>> When a path query is issued, Indeed sl=1 is returned but I see that a
>>>> rate=6 (20Gbs) is returned where I configured rate=2 (2.5 Gbs).
>>>
>>> For my best knowledge rate=2 in partition config file will be related to
>>> corresponded IPoIB multicast group for this partition, and not to
>>> PathRecord. In PathRecord you get maximum available rate on the requested
>>> path.
>>
>> I understand your comment about the relation to multicast join and not path
>> queries. However,  currently, where  there's no rule in the qos-config file
>> (or no file) that matches the path query, the SM does provide the SL
>> assigned to the partition (specified in the query) through the pkey file but
>> it doesn't do so for the Rate. So you say that for QoS = <SL, Rate>
>> assignment one should use the qos-policy file, let it be.
>
> I think Sasha is not saying "should use qos-policy file". You're
> asking about the pre QoS annex Qos implementation in OpenSM and I
> think this could be viewed as an omission (bug/feature). I think it
> could easily be changed in SA PR/MPR support.

It's the current semantics of rate just applying to the multicast
group in the partition policy file as Sasha pointed out. Unicast
traffic would be disadvantaged if using that rate. So if this were to
be done, it would need another flag for these semantics there.

Is this needed ?

-- Hal

> -- Hal
>
>> Or.
>>
>> _______________________________________________
>> general mailing list
>> general at lists.openfabrics.org
>> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>>
>> To unsubscribe, please visit
>> http://openib.org/mailman/listinfo/openib-general
>>
>



More information about the general mailing list