[Gluster-users] Change transport-type on volume from tcp to rdma, tcp

Mohammed Rafi K C rkavunga at redhat.com
Wed Jul 22 08:00:44 UTC 2015



On 07/22/2015 12:55 PM, Geoffrey Letessier wrote:
> Concerning the hang, I just saw this only once with TCP protocol but,
> actually, RDMA seems to be in cause.

If you are mounting a tcp,rdma volume using tcp protocol, all the
communication will go through the tcp connection and rdma won't come in
between client and server.

> … And, after a moment (a few minutes after having restarted my
> back-transfert of around 40TB), my volume fall down (and all my rsync
> too):
> [root at atlas ~]# df -h /mnt
> df: « /mnt »: Noeud final de transport n'est pas connecté
> df: aucun système de fichiers traité
> aka "transport endpoint is not connected »

Can you sent me the following details , if possible, ?
1) mount command used, 2) volume status 3) Client, brick logs

Regards
Rafi KC

>
> Geoffrey
>
>
> ------------------------------------------------------
> Geoffrey Letessier
> Responsable informatique & ingénieur système
> UPR 9080 - CNRS - Laboratoire de Biochimie Théorique
> Institut de Biologie Physico-Chimique
> 13, rue Pierre et Marie Curie - 75005 Paris
> Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr
> <mailto:geoffrey.letessier at ibpc.fr>
>
> Le 22 juil. 2015 à 09:17, Geoffrey Letessier
> <geoffrey.letessier at cnrs.fr <mailto:geoffrey.letessier at cnrs.fr>> a écrit :
>
>> Hi Rafi,
>>
>> It’s what I do. But I note particularly this kind of trouble when I
>> mount my volumes manually.
>>
>> In addition, when I changed my transport-type from tcp or rdma to
>> tcp,rdma, I have had to restart my volume in order they can took effect. 
>>
>> I wonder if these trouble are not due to RDMA protocol… because it
>> looks like more stable with TCP one.
>>
>> Another idea?
>> Thanks for replying and by advance,
>> Geoffrey
>> ------------------------------------------------------
>> Geoffrey Letessier
>> Responsable informatique & ingénieur système
>> UPR 9080 - CNRS - Laboratoire de Biochimie Théorique
>> Institut de Biologie Physico-Chimique
>> 13, rue Pierre et Marie Curie - 75005 Paris
>> Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr
>> <mailto:geoffrey.letessier at ibpc.fr>
>>
>> Le 22 juil. 2015 à 07:33, Mohammed Rafi K C <rkavunga at redhat.com
>> <mailto:rkavunga at redhat.com>> a écrit :
>>
>>>
>>>
>>> On 07/22/2015 04:51 AM, Geoffrey Letessier wrote:
>>>> Hi Niels,
>>>>
>>>> Thanks for replying. 
>>>>
>>>> In fact, after having checked the log, I've discovered GlusterFS
>>>> tried to connect a brick with a TCP (or RDMA) port allocated to
>>>> another volume… (bug?)
>>>> For example, here is a extract of my workdir.log file :
>>>> [2015-07-21 21:34:01.820188] E
>>>> [socket.c:2332:socket_connect_finish] 0-vol_workdir_amd-client-0:
>>>> connection to 10.0.4.1:49161 failed (Connexion refusée)
>>>> [2015-07-21 21:34:01.822563] E
>>>> [socket.c:2332:socket_connect_finish] 0-vol_workdir_amd-client-2:
>>>> connection to 10.0.4.1:49162 failed (Connexion refusée)
>>>>
>>>> But the 2 ports (49161 and 49162) concerned only my vol_home
>>>> volume, not the vol_workdir_amd one.
>>>>
>>>> Now, after having restart all glusterd synchronously (pdsh -w
>>>> cl-storage[1-4] service glusterd restart), all seems to be back
>>>> into a normal situation (size, write permission, etc.)
>>>>
>>>> But, a few minutes later, i note a strange thing I notice since
>>>> i’ve upgraded my cluster storage from 3.5.3 to 3.7.2-3: when I try
>>>> to mount some volume (particularly my vol_shared volume (replicated
>>>> volume)) my system can hang… And, because I use it in my bashrc
>>>> file for my environment modules, i need to restart my node. Idem if
>>>> I try to do a DF on my mounted volume (if it doesn’t hang during
>>>> the mount).
>>>>
>>>> With TCP transport-type, the situation seems to be more stable..
>>>>
>>>> In addition: If I restart a storage node, I can’t use Gluster CLI
>>>> (it also hang).
>>>>
>>>> Do you have an idea?
>>>
>>> Are you using bash script to start/mount the volume ? If so, add a
>>> sleep after volume start and mount, to allow all the process to
>>> start properly. Because RDMA protocol will take some time to init
>>> the resources.
>>>
>>> Regards
>>> Rafi KC
>>>
>>>
>>>
>>>>
>>>> One more time, thanks a lot for your help,
>>>> Geoffrey
>>>>
>>>> ------------------------------------------------------
>>>> Geoffrey Letessier
>>>> Responsable informatique & ingénieur système
>>>> UPR 9080 - CNRS - Laboratoire de Biochimie Théorique
>>>> Institut de Biologie Physico-Chimique
>>>> 13, rue Pierre et Marie Curie - 75005 Paris
>>>> Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr
>>>> <mailto:geoffrey.letessier at ibpc.fr>
>>>>
>>>> Le 21 juil. 2015 à 23:49, Niels de Vos <ndevos at redhat.com
>>>> <mailto:ndevos at redhat.com>> a écrit :
>>>>
>>>>> On Tue, Jul 21, 2015 at 11:20:20PM +0200, Geoffrey Letessier wrote:
>>>>>> Hello Soumya, Hello everybody,
>>>>>>
>>>>>> network.ping-timeout was set to 42 seconds. I set it to 0 but no
>>>>>> difference. The problem was, after having re-set le transport-type to
>>>>>> rdma,tcp some brick down after a few minutes.. Despite of restarting
>>>>>> volumes, after a few minutes, some [other/different] bricks down
>>>>>> again.
>>>>>
>>>>> I'm not sure how if the ping-timeout is differently handled when
>>>>> RDMA is
>>>>> used. Adding two of the guys that know RDMA well on CC.
>>>>>
>>>>>> Now, after re-creation of my volume, bricks keep alive but,
>>>>>> oddly, i’m
>>>>>> not able to write on my volume. In addition, I defined a distributed
>>>>>> volume with 2 servers, 4 bricks of 250GB each and my final volume
>>>>>> seems to be only sized to 500GB… It’s amazing..
>>>>>
>>>>> As seen further below, the 500GB volume is caused by two unreachable
>>>>> bricks. When the bricks are not reachable, the size of the bricks can
>>>>> not be detected by the client and therefore 2x 250 GB is missing.
>>>>>
>>>>> It is unclear to me why writing to a pure distributed volume
>>>>> fails. When
>>>>> a brick is not reachable, and the file should be created there, it
>>>>> would normally get created on an other brick. When the brick that
>>>>> should
>>>>> have the file gets online, and a new lookup for the file is done, a so
>>>>> called "link file" is created, which points to the file on the other
>>>>> brick. I guess the failure has to do with the connection issues, and I
>>>>> would suggest to get that solved first.
>>>>>
>>>>> HTH,
>>>>> Niels
>>>>>
>>>>>
>>>>>> Here you can find some information:
>>>>>> # gluster volume status vol_workdir_amd
>>>>>> Status of volume: vol_workdir_amd
>>>>>> Gluster process                             TCP Port  RDMA Port
>>>>>>  Online  Pid
>>>>>> ------------------------------------------------------------------------------
>>>>>> Brick ib-storage1:/export/brick_workdir/bri
>>>>>> ck1/data                                    49185     49186
>>>>>>      Y       23098
>>>>>> Brick ib-storage3:/export/brick_workdir/bri
>>>>>> ck1/data                                    49158     49159
>>>>>>      Y       3886
>>>>>> Brick ib-storage1:/export/brick_workdir/bri
>>>>>> ck2/data                                    49187     49188
>>>>>>      Y       23117
>>>>>> Brick ib-storage3:/export/brick_workdir/bri
>>>>>> ck2/data                                    49160     49161
>>>>>>      Y       3905
>>>>>>
>>>>>> # gluster volume info vol_workdir_amd
>>>>>>
>>>>>> Volume Name: vol_workdir_amd
>>>>>> Type: Distribute
>>>>>> Volume ID: 087d26ea-c6df-4cbe-94af-ecd87b59aedb
>>>>>> Status: Started
>>>>>> Number of Bricks: 4
>>>>>> Transport-type: tcp,rdma
>>>>>> Bricks:
>>>>>> Brick1: ib-storage1:/export/brick_workdir/brick1/data
>>>>>> Brick2: ib-storage3:/export/brick_workdir/brick1/data
>>>>>> Brick3: ib-storage1:/export/brick_workdir/brick2/data
>>>>>> Brick4: ib-storage3:/export/brick_workdir/brick2/data
>>>>>> Options Reconfigured:
>>>>>> performance.readdir-ahead: on
>>>>>>
>>>>>> # pdsh -w storage[1,3] df -h /export/brick_workdir/brick{1,2}
>>>>>> storage3: Filesystem            Size  Used Avail Use% Mounted on
>>>>>> storage3: /dev/mapper/st--block1-blk1--workdir
>>>>>> storage3:                       250G   34M  250G   1%
>>>>>> /export/brick_workdir/brick1
>>>>>> storage3: /dev/mapper/st--block2-blk2--workdir
>>>>>> storage3:                       250G   34M  250G   1%
>>>>>> /export/brick_workdir/brick2
>>>>>> storage1: Filesystem            Size  Used Avail Use% Mounted on
>>>>>> storage1: /dev/mapper/st--block1-blk1--workdir
>>>>>> storage1:                       250G   33M  250G   1%
>>>>>> /export/brick_workdir/brick1
>>>>>> storage1: /dev/mapper/st--block2-blk2--workdir
>>>>>> storage1:                       250G   33M  250G   1%
>>>>>> /export/brick_workdir/brick2
>>>>>>
>>>>>> # df -h /workdir/
>>>>>> Filesystem            Size  Used Avail Use% Mounted on
>>>>>> localhost:vol_workdir_amd.rdma
>>>>>>                      500G   67M  500G   1% /workdir
>>>>>>
>>>>>> # touch /workdir/test
>>>>>> touch: impossible de faire un touch « /workdir/test »: Aucun
>>>>>> fichier ou dossier de ce type
>>>>>>
>>>>>> # tail -30l /var/log/glusterfs/workdir.log
>>>>>> Host Unreachable, Check your connection with IPoIB
>>>>>> [2015-07-21 21:10:33.927673] W
>>>>>> [rdma.c:1263:gf_rdma_cm_event_handler]
>>>>>> 0-vol_workdir_amd-client-2: cma event RDMA_CM_EVENT_REJECTED,
>>>>>> error 8 (me:10.0.4.1:1020 peer:10.0.4.1:49174)
>>>>>> Host Unreachable, Check your connection with IPoIB
>>>>>> [2015-07-21 21:10:37.877231] I
>>>>>> [rpc-clnt.c:1819:rpc_clnt_reconfig] 0-vol_workdir_amd-client-0:
>>>>>> changing port to 49173 (from 0)
>>>>>> [2015-07-21 21:10:37.880556] I
>>>>>> [rpc-clnt.c:1819:rpc_clnt_reconfig] 0-vol_workdir_amd-client-2:
>>>>>> changing port to 49174 (from 0)
>>>>>> [2015-07-21 21:10:37.914661] W
>>>>>> [rdma.c:1263:gf_rdma_cm_event_handler]
>>>>>> 0-vol_workdir_amd-client-0: cma event RDMA_CM_EVENT_REJECTED,
>>>>>> error 8 (me:10.0.4.1:1021 peer:10.0.4.1:49173)
>>>>>> Host Unreachable, Check your connection with IPoIB
>>>>>> [2015-07-21 21:10:37.923535] W
>>>>>> [rdma.c:1263:gf_rdma_cm_event_handler]
>>>>>> 0-vol_workdir_amd-client-2: cma event RDMA_CM_EVENT_REJECTED,
>>>>>> error 8 (me:10.0.4.1:1020 peer:10.0.4.1:49174)
>>>>>> Host Unreachable, Check your connection with IPoIB
>>>>>> [2015-07-21 21:10:41.883925] I
>>>>>> [rpc-clnt.c:1819:rpc_clnt_reconfig] 0-vol_workdir_amd-client-0:
>>>>>> changing port to 49173 (from 0)
>>>>>> [2015-07-21 21:10:41.887085] I
>>>>>> [rpc-clnt.c:1819:rpc_clnt_reconfig] 0-vol_workdir_amd-client-2:
>>>>>> changing port to 49174 (from 0)
>>>>>> [2015-07-21 21:10:41.919394] W
>>>>>> [rdma.c:1263:gf_rdma_cm_event_handler]
>>>>>> 0-vol_workdir_amd-client-0: cma event RDMA_CM_EVENT_REJECTED,
>>>>>> error 8 (me:10.0.4.1:1021 peer:10.0.4.1:49173)
>>>>>> Host Unreachable, Check your connection with IPoIB
>>>>>> [2015-07-21 21:10:41.932622] W
>>>>>> [rdma.c:1263:gf_rdma_cm_event_handler]
>>>>>> 0-vol_workdir_amd-client-2: cma event RDMA_CM_EVENT_REJECTED,
>>>>>> error 8 (me:10.0.4.1:1020 peer:10.0.4.1:49174)
>>>>>> Host Unreachable, Check your connection with IPoIB
>>>>>> [2015-07-21 21:10:44.682636] W
>>>>>> [dht-layout.c:189:dht_layout_search] 0-vol_workdir_amd-dht: no
>>>>>> subvolume for hash (value) = 1072520554
>>>>>> [2015-07-21 21:10:44.682947] W
>>>>>> [dht-layout.c:189:dht_layout_search] 0-vol_workdir_amd-dht: no
>>>>>> subvolume for hash (value) = 1072520554
>>>>>> [2015-07-21 21:10:44.683240] W
>>>>>> [dht-layout.c:189:dht_layout_search] 0-vol_workdir_amd-dht: no
>>>>>> subvolume for hash (value) = 1072520554
>>>>>> [2015-07-21 21:10:44.683472] W
>>>>>> [dht-diskusage.c:48:dht_du_info_cbk] 0-vol_workdir_amd-dht:
>>>>>> failed to get disk info from vol_workdir_amd-client-0
>>>>>> [2015-07-21 21:10:44.683506] W
>>>>>> [dht-diskusage.c:48:dht_du_info_cbk] 0-vol_workdir_amd-dht:
>>>>>> failed to get disk info from vol_workdir_amd-client-2
>>>>>> [2015-07-21 21:10:44.683532] W
>>>>>> [dht-layout.c:189:dht_layout_search] 0-vol_workdir_amd-dht: no
>>>>>> subvolume for hash (value) = 1072520554
>>>>>> [2015-07-21 21:10:44.683551] W
>>>>>> [fuse-bridge.c:1970:fuse_create_cbk] 0-glusterfs-fuse: 18: /test
>>>>>> => -1 (Aucun fichier ou dossier de ce type)
>>>>>> [2015-07-21 21:10:44.683619] W
>>>>>> [dht-layout.c:189:dht_layout_search] 0-vol_workdir_amd-dht: no
>>>>>> subvolume for hash (value) = 1072520554
>>>>>> [2015-07-21 21:10:44.683846] W
>>>>>> [dht-layout.c:189:dht_layout_search] 0-vol_workdir_amd-dht: no
>>>>>> subvolume for hash (value) = 1072520554
>>>>>> [2015-07-21 21:10:45.886807] I
>>>>>> [rpc-clnt.c:1819:rpc_clnt_reconfig] 0-vol_workdir_amd-client-0:
>>>>>> changing port to 49173 (from 0)
>>>>>> [2015-07-21 21:10:45.893059] I
>>>>>> [rpc-clnt.c:1819:rpc_clnt_reconfig] 0-vol_workdir_amd-client-2:
>>>>>> changing port to 49174 (from 0)
>>>>>> [2015-07-21 21:10:45.920434] W
>>>>>> [rdma.c:1263:gf_rdma_cm_event_handler]
>>>>>> 0-vol_workdir_amd-client-0: cma event RDMA_CM_EVENT_REJECTED,
>>>>>> error 8 (me:10.0.4.1:1021 peer:10.0.4.1:49173)
>>>>>> Host Unreachable, Check your connection with IPoIB
>>>>>> [2015-07-21 21:10:45.925292] W
>>>>>> [rdma.c:1263:gf_rdma_cm_event_handler]
>>>>>> 0-vol_workdir_amd-client-2: cma event RDMA_CM_EVENT_REJECTED,
>>>>>> error 8 (me:10.0.4.1:1020 peer:10.0.4.1:49174)
>>>>>> Host Unreachable, Check your connection with IPoIB
>>>>>>
>>>>>> I use GlusterFS in production since around 3 years without any block
>>>>>> problem but now the situation is awesome since more than 3 weeks…
>>>>>> Indeed, our production are down since roughly 3.5 weeks (with a lot
>>>>>> and different problems with GlusterFS v3.5.3 and now with
>>>>>> 3.7.2-3) and
>>>>>> i need to restart it…
>>>>>>
>>>>>> Thanks in advance,
>>>>>> Geoffrey
>>>>>> ------------------------------------------------------
>>>>>> Geoffrey Letessier
>>>>>> Responsable informatique & ingénieur système
>>>>>> UPR 9080 - CNRS - Laboratoire de Biochimie Théorique
>>>>>> Institut de Biologie Physico-Chimique
>>>>>> 13, rue Pierre et Marie Curie - 75005 Paris
>>>>>> Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr
>>>>>> <mailto:geoffrey.letessier at ibpc.fr>
>>>>>>
>>>>>> Le 21 juil. 2015 à 19:36, Soumya Koduri <skoduri at redhat.com
>>>>>> <mailto:skoduri at redhat.com>> a écrit :
>>>>>>
>>>>>>> From the following errors,
>>>>>>>
>>>>>>> [2015-07-21 14:36:30.495321] I [MSGID: 114020]
>>>>>>> [client.c:2118:notify] 0-vol_shared-client-0: parent translators
>>>>>>> are ready, attempting connect on transport
>>>>>>> [2015-07-21 14:36:30.498989] W [socket.c:923:__socket_keepalive]
>>>>>>> 0-socket: failed to set TCP_USER_TIMEOUT 0 on socket 12,
>>>>>>> Protocole non disponible
>>>>>>> [2015-07-21 14:36:30.499004] E [socket.c:3015:socket_connect]
>>>>>>> 0-vol_shared-client-0: Failed to set keep-alive: Protocole non
>>>>>>> disponible
>>>>>>>
>>>>>>> looks like setting TCP_USER_TIMEOUT value to 0 on the socket
>>>>>>> failed with error (IIUC) "Protocol not available".
>>>>>>> Could you check if 'network.ping-timeout' is set to zero for
>>>>>>> that volume using 'gluster volume info'? Anyways from the code
>>>>>>> looks like 'TCP_USER_TIMEOUT' can take value zero. Not sure why
>>>>>>> it has failed.
>>>>>>>
>>>>>>> Niels, any thoughts?
>>>>>>>
>>>>>>> Thanks,
>>>>>>> Soumya
>>>>>>>
>>>>>>> On 07/21/2015 08:15 PM, Geoffrey Letessier wrote:
>>>>>>>> [2015-07-21 14:36:30.495321] I [MSGID: 114020]
>>>>>>>> [client.c:2118:notify]
>>>>>>>> 0-vol_shared-client-0: parent translators are ready, attempting
>>>>>>>> connect
>>>>>>>> on transport
>>>>>>>> [2015-07-21 14:36:30.498989] W [socket.c:923:__socket_keepalive]
>>>>>>>> 0-socket: failed to set TCP_USER_TIMEOUT 0 on socket 12,
>>>>>>>> Protocole non
>>>>>>>> disponible
>>>>>>>> [2015-07-21 14:36:30.499004] E [socket.c:3015:socket_connect]
>>>>>>>> 0-vol_shared-client-0: Failed to set keep-alive: Protocole non
>>>>>>>> disponible
>>>>>>
>>>>
>>>
>>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20150722/1bd7eea5/attachment.html>


More information about the Gluster-users mailing list