[ofa-general] Re: [RFC] OpenSM and IPv6 Scalability Proposal

[Slava Strebkov]

-----Original Message-----
From: Hal Rosenstock [mailto:hal.rosenstock at gmail.com] 
Sent: Friday, May 08, 2009 4:58 PM
To: Slava Strebkov
Cc: general at lists.openfabrics.org
Subject: Re: [ofa-general] Re: [RFC] OpenSM and IPv6 Scalability Proposal

On Wed, May 6, 2009 at 6:24 AM, Slava Strebkov <slavas at voltaire.com> wrote:
>
> In addition to the original proposal we suggest allocating special MLID
> for the following MGIDs:
>  1. FF12401bxxxx000000000000FFFFFFFF - All Nodes
>  2. FF12401bxxxx00000000000000000001 - All hosts
>  3. FF12401bffff0000000000000000004d  - all Gateways
>  4. FF12401bxxxx00000000000000000002 - all routers
>  5. FF12601bABCD000000000001ffxxxxxx - IPv6 SNM

It turns out that collapsing multicast groups across PKeys on a single
MLID may not be such a good idea unless partition enforcement
enforcement by switches is disabled. There should be different modes
of collapsing based on this based on whether this is enabled or not.

> For all other cases we suggest that same MLID will be assigned to
> different MGIDs if:
>  1. They share the same P Key
>  2. Same signature - for IPoIB only
>  3. Same LSB bits - bitmask configurable by user (default  10 bits)
>        for example, the following are the same:
>        MGID1:  FF12401bABCD000000000000xxxxx755
>        MGID2:  FF12401bABCD000000000000yyyyyB55

Jason's approach to this was in a thread entitled "IPv6 and IPoIB
scalability issue":
http://lists.openfabrics.org/pipermail/general/2006-November/029621.html
in which he proposed an MGID range (MGID/prefix syntax) for collapsing
IPv6 SNM groups. Additionally, there was the potential to distribute
the matched groups across some number of MLIDs. See also thread "[RFC]
OpenSM and IPv6 Scalability Proposal":
http://lists.openfabrics.org/pipermail/general/2008-June/051226.html

>  Implementation.
>  Since there will be many mgroups shared same mlid, mlid-array entry
> will contain
>  fleximap holding mgroups.
>  Searching of mgroup will be performed by mlid (index in the array) and
> mgid -
>  key in the fleximap.

Sasha proposed using an array rather than fleximap for this:
http://lists.openfabrics.org/pipermail/general/2008-June/051525.html

We propose MLID -indexed array, but instead of list of pointers to multicast groups, there will be fleximap sorted by MGID. This is faster than simple list.
Slava

-- Hal

>
>
>  Slava Strebkov
> _______________________________________________
> 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
>