crossbow discuss - Sep 2009 - QoS Support in Solaris/Crossbow

>
>
> I think I am correct that HW Tx/Rx rings <=> packet/flow mapping
> is not fully controllable/manipulated by flowadm of Crossbow.  But,
> do we know how many HW classification features of Oplin/Niantic
> are utilized by Crossbow? For example (just as an example), ToS/dscp
> (or diffsrv) classification can utilize NIC HW or not as of today?
> How about future Crossbow?
Jan-Lung, as far as I know the hardware assisted L3/L4 flow 
classification will be supported in Crossbow phase II. Right now, one 
can make use of "dsfield" attribute while creating a flow to state the
bits in the IP header. But the flow classification is done by software. 
Cc''ing crossbow-discuss. Please correct us if that''s not the
case.
>
> BTW, does Crossbow currently bind Tx and Rx HW rings together?
> Could you share how given that Crossbow tx classification has to
> be very intelligent to work with NIC HW-specific classification engine
> to bind them together..... or, there is a way of flow tagging <=> HW
ring
> mapping? (I guess this is more likely the way).
I don''t  see Tx and Rx hardware rings are bound together for VNICs on
my
setup. Crossbow Phase II will support dynamic ring grouping - specify 
number of rings during or after VNIC creation. But I guess this is only 
on Tx side. If this is for both Tx and Rx hardware rings, then you can 
get the flow to hardware ring mapping by creating flows on top of the VNIC.
Jing
>
> Jan-Lung
>
>

Jing Zhang wrote:
> 
>>
>>
>> I think I am correct that HW Tx/Rx rings <=> packet/flow mapping
>> is not fully controllable/manipulated by flowadm of Crossbow.  But,
>> do we know how many HW classification features of Oplin/Niantic
>> are utilized by Crossbow? For example (just as an example), ToS/dscp
>> (or diffsrv) classification can utilize NIC HW or not as of today?
>> How about future Crossbow?
> Jan-Lung, as far as I know the hardware assisted L3/L4 flow 
> classification will be supported in Crossbow phase II. Right now, one 
> can make use of "dsfield" attribute while creating a flow to
state the
> bits in the IP header. But the flow classification is done by software. 
> Cc''ing crossbow-discuss. Please correct us if that''s not
the case.
That is correct. Flowadm currently uses S/W classification. We are using
H/W classification for only L2 classifications. Crossbow phase 2 plans
to use H/W classification. We also do S/W classification on DSCP
fields today (although outbound marking is missing). The outbound
marking and H/W support (where available) will also come in phase 2.
>>
>> BTW, does Crossbow currently bind Tx and Rx HW rings together?
>> Could you share how given that Crossbow tx classification has to
>> be very intelligent to work with NIC HW-specific classification engine
>> to bind them together..... or, there is a way of flow tagging <=>
HW ring
>> mapping? (I guess this is more likely the way).
> I don''t  see Tx and Rx hardware rings are bound together for VNICs
on my
> setup. Crossbow Phase II will support dynamic ring grouping - specify 
> number of rings during or after VNIC creation. But I guess this is only 
> on Tx side. If this is for both Tx and Rx hardware rings, then you can 
> get the flow to hardware ring mapping by creating flows on top of the VNIC.
> Jing
The current code has asymmetric behavior for Tx. We already have
prototype code that make Tx and Rx symmetric and binds Rx and Tx
together on per link basis. This code will mostly get in phase 1.3
targeting next OpenSolaris release. Let us know if you have some
specific needs in either of these areas.

Cheers,
Sunay