[openib-general] IPv6 and IPoIB scalability issue

Hal Rosenstock halr at voltaire.com
Thu Nov 30 14:29:16 PST 2006


On Thu, 2006-11-30 at 17:26, Todd Rimmer wrote:
> There is a potential limitation in IB which can affect IPv6 operation on
> larger clusters.
> 
> IPV6 defines that each node will have a Solicited Node Multicast
> address.  This address is unique per node and is constructed from the
> IPV6 unicast address of the node.  (see RFC 2373 for more details).
> 
> IP over IB defines that IPV6 multicast addresses map to IB multicast
> GIDs in a one to one manner.
> 
> IB defines a multicast address space limit of 4095 LIDs.

actually it is 16K-1

> Popular IB switches have a smaller limit, typically 1024 multicast
> forwarding table entries.

This is current practical limit.

> This means that if IPv6 is enabled, a cluster of ~1024 nodes may run out
> of multicast entries in the switches and may encounter problems when
> running IPv6 traffic. (For example if the Solicited Node Multicast
> addresses consume all the multicast switch entries, other application
> specific multicast groups or worse yet permanent multicast addresses
> could fail to be created).
> 
> Proposed solution:
> - add an IPoIB configuration parameter.  This parameter could redirect
> the Solicited Node Multicast traffic to the IPv6 All Nodes multicast
> address (IB GID 0xff01601B.....0000001)
> - on clusters near 1024 nodes (there are a few other standard multicast
> addresses not to mention application specific multicast for IPv4 and
> IPv6) this parameter would be required to be enabled.
> 
> While this means that Solicited Node Multicast becomes a much larger
> scope multicast, it will greatly reduce the Multicast member record
> stress on the SA and switches.  In general the IP stack will filter
> undesired inbound multicast packets, so applications will continue to
> function properly.
> 
> While not ideal, this approach solves the problem easily.  Alternatives
> based on multicast member record join results or joining/creating the
> group only if really needed are possible but can be more complex and
> have subtle complications and issues.
> 
> Being configurable would allow clusters where efficiency for Solicited
> Node Multicast was important to choose the appropriate trade-off in
> operation.
> 
> Thoughts?

MGIDs are different from MLIDs. Multiple MGIDs can be mapped onto a
single MLID if the characteristics are the same. Is that the case for
the IPv6 groups ?

The problem I have seen is that the link local ones chew up a MLID on
those SMs which use a MLID per MGID.

-- Hal

> Todd Rimmer
> Chief Architect         System Interconnect Group, QLogic
> Voice: 610-233-4852     Fax: 610-233-4777
> Todd.Rimmer at QLogic.com  www.QLogic.com
>  
> 
> > -----Original Message-----
> > From: openib-general-bounces at openib.org [mailto:openib-general-
> > bounces at openib.org] On Behalf Of Sean Hubbell
> > Sent: Wednesday, November 29, 2006 2:11 PM
> > To: Eitan Zahavi
> > Cc: openib-general at openib.org
> > Subject: Re: [openib-general] SDP Protocol Issue Was: Configuration of
> sdp
> > 
> > Sounds like the problem I am having as I downloaded the lastest and
> > tried but still get the error 97 which is protocol family. I'll
> upgrade
> > the latest ib modules, but to do that I have to upgrade my kernel
> first...
> > 
> > Thanks...
> > 
> > Sean
> > 
> > Eitan Zahavi wrote:
> > > Hi Sean,
> > >
> > > Now I remember that if you go back enough with SDP (I think MST
> should
> > > know how far) the Address Family in the address given to SDP used to
> be
> > > required to be of type AF_INET_SDP. The new libsdp code does not do
> that
> > > and works with the new SDP requirement to only create the socket
> with
> > > AF_INET_SDP and later use AF_INET for the address (used for
> > > bind/connect).
> > >
> > > So you should try and move to newest SDP too.
> > >
> > > Eitan Zahavi
> > > Senior Engineering Director, Software Architect
> > > Mellanox Technologies LTD
> > > Tel:+972-4-9097208
> > > Fax:+972-4-9593245
> > > P.O. Box 586 Yokneam 20692 ISRAEL
> > >
> > >
> > >
> > >> -----Original Message-----
> > >> From: Sean Hubbell [mailto:shubbell at dbresearch.net]
> > >> Sent: Wednesday, November 29, 2006 7:53 PM
> > >> To: Eitan Zahavi; openib-general at openib.org
> > >> Subject: Re: [openib-general] SDP Protocol Issue Was: Configuration
> of
> > >>
> > > sdp
> > >
> > >> Eitan Zahavi wrote:
> > >>
> > >>> Sean Hubbell wrote:
> > >>>
> > >>>> Ok, after futher inverstigation, here is what I have found:
> > >>>>
> > >>>> socket failed: Address family not supported by protocol
> > >>>>
> > >>>>
> > >>> Do yo u have sdp loaded. Please try:
> > >>> lsmod | grep ib_sdp
> > >>>
> > >> Yes.
> > >>
> > >>>> I am assuming when sdp "exchanges" the AF_INET sockets to AF_SDP
> > >>>> sockets that the current code that I am running has a issue or I
> > >>>> don't know how to configure things...
> > >>>>
> > >>>> I have running libsdp v. 0.9.0 x86_64 as the rpm from downloaded
> > >>>>
> > > from
> > >
> > >>>> mirror.centos.org running kernel 2.6.9-42.0.3.plus.c4smp
> > >>>>
> > >>>>
> > >>>>
> > >>> I would recommend upgrading libsdp only to the one available from
> > >>>
> > > the
> > >
> > >>> git tree at git clone
> > >>> git://staging.openfabrics.org/git/~eitan/libsdp.git
> > >>>
> > >>> or still fom SVN:
> > >>> https://openib.org/svn/gen2/trunk/src/userspace/libsdp
> > >>>
> > >>>
> > >>>> Also, my /etc/libsdp.conf file has the following:
> > >>>>
> > >>>> log min-level 1 destination syslog
> > >>>> #use both server * *:*
> > >>>> #use both client * *:*
> > >>>> use sdp client * 10.10.0.0/16:*
> > >>>> use sdp server * 10.10.0.0/16:*
> > >>>>
> > >>>> and seems to break all of my TCP connections, regardless if the
> > >>>> interface is 10.10.* Is my configuration correct?
> > >>>>
> > >>>>
> > >>> The "both" would have been better in preserving TCP based services
> > >>>
> > > as
> > >
> > >>> it falls back to use TCP if SDP does not work.
> > >>>
> > >> I tried that originally and received the same results (things did
> not
> > >>
> > > fall back to
> > >
> > >> using tcp), so I attempted to isolate this to my infiniband
> > >>
> > > connection... I'll
> > >
> > >> change it to both.
> > >>
> > >>
> > >>>> What would you recommend my next step to be? Upgrading my kernel
> as
> > >>>> well as libsdp to the latest? Also, currently without using sdp I
> > >>>>
> > > am
> > >
> > >>>> getting 941 MBps, what do you get when you run using sdp, Karun
> > >>>> mentioned that he is getting double the performance, which is
> where
> > >>>>
> > > I
> > >
> > >>>> am at; needing better performance from my apps?
> > >>>>
> > >>>>
> > >>> Depends on your apps. If you have many connections of large
> messages
> > >>> SDP is a clear win.
> > >>>
> > >> Yes, so it looks like this would be a valid upgrade. I'll try the
> git
> > >>
> > > tree.
> > >
> > >> Thanks Eitan,
> > >>
> > >> Sean
> > >>
> > >
> > >
> > >
> > 
> > 
> > _______________________________________________
> > openib-general mailing list
> > openib-general at openib.org
> > http://openib.org/mailman/listinfo/openib-general
> > 
> > To unsubscribe, please visit
> http://openib.org/mailman/listinfo/openib-
> > general
> 
> 
> _______________________________________________
> openib-general mailing list
> openib-general at openib.org
> http://openib.org/mailman/listinfo/openib-general
> 
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
> 





More information about the general mailing list