[Users] Weird IPoIB issue
Robert LeBlanc
robert_leblanc at BYU.EDU
Mon Oct 28 11:16:24 PDT 2013
I was just wondering about that. It seems reasonable that the broadcast
traffic would go over multicast, but effectively channels would be created
for node to node communication, otherwise the entire multicast group would
be limited to 10 Gbps (in this instance) for the whole group. That doesn't
scale very well.
The things I've read about IPoIB performance tuning seems pretty vague, and
the changes most people recommend seem to be already in place on the
systems I'm using. Some people said, try using a newer version of Ubuntu,
but ultimately, I have very little control over VMware. Once I can get the
Linux machines to communicate IPoIB between the racks and blades, then I'm
going to turn my attention over to performance optimization. It doesn't
seem to make much sense to spend time there when it is not working at all
for most machines.
I've done ibtracert between the two nodes, is that what you mean by walking
the route?
[root at desxi003 ~]# ibtracert -m 0xc000 0x2d 0x37
>From ca 0xf04da2909778e7d0 port 1 lid 45-45 "localhost HCA-1"
[1] -> switch 0x2c90200448ec8[17] lid 51 "Infiniscale-IV Mellanox
Technologies"
[18] -> ca 0xf04da2909778e714[1] lid 55 "localhost HCA-1"
To ca 0xf04da2909778e714 port 1 lid 55-55 "localhost HCA-1"
[root at desxi004 ~]# ibtracert -m 0xc000 0x37 0x2d
>From ca 0xf04da2909778e714 port 1 lid 55-55 "localhost HCA-1"
[1] -> switch 0x2c90200448ec8[18] lid 51 "Infiniscale-IV Mellanox
Technologies"
[17] -> ca 0xf04da2909778e7d0[1] lid 45 "localhost HCA-1"
To ca 0xf04da2909778e7d0 port 1 lid 45-45 "localhost HCA-1"
As you can see, the route is on the same switch, the blades are right next
to each other.
Robert LeBlanc
OIT Infrastructure & Virtualization Engineer
Brigham Young University
On Mon, Oct 28, 2013 at 12:05 PM, Hal Rosenstock
<hal.rosenstock at gmail.com>wrote:
> Which mystery is explained ? The 10 Gbps is a multicast only limit and
> does not apply to unicast. The BW limitation you're seeing is due to other
> factors. There's been much written about IPoIB performance.
>
> If all the MC members are joined and routed, then the IPoIB connectivity
> issue is some other issue. Are you sure this is the case ? Did you walk the
> route between 2 nodes where you have a connectivity issue ?
>
>
> On Mon, Oct 28, 2013 at 1:58 PM, Robert LeBlanc <robert_leblanc at byu.edu>wrote:
>
>> Well, that explains one mystery, now I need to figure out why it seems
>> the Dell blades are not passing the traffic.
>>
>>
>> Robert LeBlanc
>> OIT Infrastructure & Virtualization Engineer
>> Brigham Young University
>>
>>
>> On Mon, Oct 28, 2013 at 11:51 AM, Hal Rosenstock <
>> hal.rosenstock at gmail.com> wrote:
>>
>>> Yes, that's the IPoIB IPv4 broadcast group for the default (0xffff)
>>> partition. 0x80 part of mtu and rate just means "is equal to". mtu 0x04 is
>>> 2K (2048) and rate 0x3 is 10 Gb/sec. These are indeed the defaults.
>>>
>>>
>>> On Mon, Oct 28, 2013 at 1:45 PM, Robert LeBlanc <robert_leblanc at byu.edu>wrote:
>>>
>>>> The info for that MGID is:
>>>> MCMemberRecord group dump:
>>>> MGID....................ff12:401b:ffff::ffff:ffff
>>>> Mlid....................0xC000
>>>> Mtu.....................0x84
>>>> pkey....................0xFFFF
>>>> Rate....................0x83
>>>> SL......................0x0
>>>>
>>>> I don't understand the MTU and Rate (130 and 131 dec). When I run iperf
>>>> between the two hosts over IPoIB in connected mode and MTU 65520. I've
>>>> tried multiple threads, but the sum is still 10 Gbps.
>>>>
>>>>
>>>> Robert LeBlanc
>>>> OIT Infrastructure & Virtualization Engineer
>>>> Brigham Young University
>>>>
>>>>
>>>> On Mon, Oct 28, 2013 at 11:40 AM, Hal Rosenstock <
>>>> hal.rosenstock at gmail.com> wrote:
>>>>
>>>>> saquery -g should show what MGID is mapped to MLID 0xc000 and the
>>>>> group parameters.
>>>>>
>>>>> When you say 10 Gbps max, is that multicast or unicast ? That limit is
>>>>> only on the multicast.
>>>>>
>>>>>
>>>>> On Mon, Oct 28, 2013 at 1:28 PM, Robert LeBlanc <
>>>>> robert_leblanc at byu.edu> wrote:
>>>>>
>>>>>> Well, that can explain why I'm only able to get 10 Gbps max from the
>>>>>> two hosts that are working.
>>>>>>
>>>>>> I have tried updn and dnup and they didn't help either. I think the
>>>>>> only thing that will help is Automatic Path Migration is it tries very hard
>>>>>> to route the alternative LIDs through different systemguids. I suspect it
>>>>>> would require re-LIDing everything which would mean an outage. I'm still
>>>>>> trying to get answers from Oracle if that is even a possibility. I've tried
>>>>>> seeding some of the algorithms with information like root nodes, etc, but
>>>>>> none of them worked better.
>>>>>>
>>>>>> The MLID 0xc000 exists and I can see all the nodes joined to the
>>>>>> group using saquery. I've checked the route using ibtracert specifying the
>>>>>> MLID. The only thing I'm not sure how to check is the group parameters.
>>>>>> What tool would I use for that?
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>>
>>>>>> Robert LeBlanc
>>>>>> OIT Infrastructure & Virtualization Engineer
>>>>>> Brigham Young University
>>>>>>
>>>>>>
>>>>>> On Mon, Oct 28, 2013 at 11:16 AM, Hal Rosenstock <
>>>>>> hal.rosenstock at gmail.com> wrote:
>>>>>>
>>>>>>> Xsigo's SM is not "straight" OpenSM. They have some proprietary
>>>>>>> enhancements and it may be based on old vintage of OpenSM. You will likely
>>>>>>> need to work with them/Oracle now on issues.
>>>>>>>
>>>>>>> Lack of a partitions file does mean default partition and default
>>>>>>> rate (10 Gbps) so from what I saw all ports had sufficient rate to join MC
>>>>>>> group.
>>>>>>>
>>>>>>> There are certain topology requirements for running various routing
>>>>>>> algorithms. Did you try updn or dnup ?
>>>>>>>
>>>>>>> The key is determining whether the IPoIB broadcast group is setup
>>>>>>> correctly. What MLID is the group built on (usually 0xc000) ? What are the
>>>>>>> group parameters (rate, MTU) ? Are all members that are running IPoIB
>>>>>>> joined ? Is the group routed to all such members ? There are
>>>>>>> infiniband-diags for all of this.
>>>>>>>
>>>>>>>
>>>>>>> On Mon, Oct 28, 2013 at 12:19 PM, Robert LeBlanc <
>>>>>>> robert_leblanc at byu.edu> wrote:
>>>>>>>
>>>>>>>> OpenSM (the SM runs on Xsigo so they manage it) is using minhop.
>>>>>>>> I've loaded the ibnetdiscover output into ibsim and run all the different
>>>>>>>> routing algorithms against it with and without scatter ports. Minhop had
>>>>>>>> 50% of our hosts running all paths through a single IS5030 switch (at least
>>>>>>>> the LIDs we need which represent Ethernet and Fibre Channel cards the hosts
>>>>>>>> should communicate with). Ftree, dor, and dfsssp failed back to minhop, the
>>>>>>>> others routed more paths through the same IS5030 in some cases increasing
>>>>>>>> our host count with single point of failure to 75%.
>>>>>>>>
>>>>>>>> As far as I can tell there is no partitions.conf file so I assume
>>>>>>>> we are using the default partition. There is an opensm.opts file, but it
>>>>>>>> only specifies logging information.
>>>>>>>> # SA database file name
>>>>>>>> sa_db_file /var/log/opensm-sa.dump
>>>>>>>>
>>>>>>>> # If TRUE causes OpenSM to dump SA database at the end of
>>>>>>>> # every light sweep, regardless of the verbosity level
>>>>>>>> sa_db_dump TRUE
>>>>>>>>
>>>>>>>> # The directory to hold the file OpenSM dumps
>>>>>>>> dump_files_dir /var/log/
>>>>>>>>
>>>>>>>> The SM node is:
>>>>>>>> xsigoa:/opt/xsigo/xsigos/current/ofed/etc# ibaddr
>>>>>>>> GID fe80::13:9702:100:979 LID start 0x1 end 0x1
>>>>>>>>
>>>>>>>> We do have Switch-X in two of the Dell m1000e chassis but the
>>>>>>>> cards, ports 17-32, are FDR10 (the switch may be straight FDR, but I'm not
>>>>>>>> 100% sure). The IS5030 are QDR which the Switch-X are connected to, the
>>>>>>>> switches in the Xsigo directors are QDR, but the Ethernet and Fibre Channel
>>>>>>>> cards are DDR. The DDR cards will not be running IPoIB (at least to my
>>>>>>>> knowledge they don't have the ability), only the hosts should be leveraging
>>>>>>>> IPoIB. I hope that clears up some of your questions. If you have more, I
>>>>>>>> will try to answer them.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Robert LeBlanc
>>>>>>>> OIT Infrastructure & Virtualization Engineer
>>>>>>>> Brigham Young University
>>>>>>>>
>>>>>>>>
>>>>>>>> On Mon, Oct 28, 2013 at 9:57 AM, Hal Rosenstock <
>>>>>>>> hal.rosenstock at gmail.com> wrote:
>>>>>>>>
>>>>>>>>> What routing algorithm is configured in OpenSM ? What does your
>>>>>>>>> partitions.conf file look like ? Which node is your OpenSM ?
>>>>>>>>>
>>>>>>>>> Also, I only see QDR and DDR links although you have Switch-X so I
>>>>>>>>> assume all FDR ports are connected to slower (QDR) devices. I don't see any
>>>>>>>>> FDR-10 ports but maybe they're also connected to QDR ports so show up as
>>>>>>>>> QDR in the topology.
>>>>>>>>>
>>>>>>>>> There are DDR CAs in Xsigo box but not sure whether or not they
>>>>>>>>> run IPoIB.
>>>>>>>>>
>>>>>>>>> -- Hal
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On Sun, Oct 27, 2013 at 9:46 PM, Robert LeBlanc <
>>>>>>>>> robert_leblanc at byu.edu> wrote:
>>>>>>>>>
>>>>>>>>>> Since you guys are amazingly helpful, I thought I would pick your
>>>>>>>>>> brains in a new problem.
>>>>>>>>>>
>>>>>>>>>> We have two Xsigo directors cross connected to four Mellanox
>>>>>>>>>> IS5030 switches. Connected to those we have four Dell m1000e chassis each
>>>>>>>>>> with two IB switches (two chassis have QDR and two have FDR10). We have 9
>>>>>>>>>> dual-port rack servers connected to the IS5030 switches. For testing
>>>>>>>>>> purposes we have an additional Dell m1000e QDR chassis connected to one
>>>>>>>>>> Xsigo director and two dual-port FDR10 rack servers connected to the other
>>>>>>>>>> Xsigo director.
>>>>>>>>>>
>>>>>>>>>> I can get IPoIB to work between the two test rack servers
>>>>>>>>>> connected to the one Xsigo director. But I can not get IPoIB to work
>>>>>>>>>> between any blades either right next to each other or to the working rack
>>>>>>>>>> servers. I'm using the same exact live CentOS ISO on all four servers. I've
>>>>>>>>>> checked opensm and the blades have joined the multicast group 0xc000
>>>>>>>>>> properly. tcpdump basically says that traffic is not leaving the blades.
>>>>>>>>>> tcpdump also shows no traffic entering the blades from the rack servers. An
>>>>>>>>>> ibtracert using 0xc000 mlid shows that routing exists between hosts.
>>>>>>>>>>
>>>>>>>>>> I've read about MulticastFDBTop=0xBFFF but I don't know how to
>>>>>>>>>> set it and I doubt it would have been set by default.
>>>>>>>>>>
>>>>>>>>>> Anyone have some ideas on troubleshooting steps to try? I think
>>>>>>>>>> Google is tired of me asking questions about it.
>>>>>>>>>>
>>>>>>>>>> Thanks,
>>>>>>>>>>
>>>>>>>>>> Robert LeBlanc
>>>>>>>>>> OIT Infrastructure & Virtualization Engineer
>>>>>>>>>> Brigham Young University
>>>>>>>>>>
>>>>>>>>>> _______________________________________________
>>>>>>>>>> Users mailing list
>>>>>>>>>> Users at lists.openfabrics.org
>>>>>>>>>> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/users
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/users/attachments/20131028/4cefc7ee/attachment.html>
More information about the Users
mailing list