On Thu, Dec 8, 2016 at 4:37 PM, Micha Ober <mich...@gmail.com> wrote:
> Hi Rafi, > > thank you for your support. It is greatly appreciated. > > Just some more thoughts from my side: > > There have been no reports from other users in *this* thread until now, > but I have found at least one user with a very simiar problem in an older > thread: > > https://www.gluster.org/pipermail/gluster-users/2014-November/019637.html > > He is also reporting disconnects with no apparent reasons, althogh his > setup is a bit more complicated, also involving a firewall. In our setup, > all servers/clients are connected via 1 GbE with no firewall or anything > that might block/throttle traffic. Also, we are using exactly the same > software versions on all nodes. > > > I can also find some reports in the bugtracker when searching for > "rpc_client_ping_timer_expired" and "rpc_clnt_ping_timer_expired" (looks > like spelling changed during versions). > > https://bugzilla.redhat.com/show_bug.cgi?id=1096729 > Just FYI, this is a different issue, here GlusterD fails to handle the volume of incoming requests on time since MT-epoll is not enabled here. > > https://bugzilla.redhat.com/show_bug.cgi?id=1370683 > > But both reports involve large traffic/load on the bricks/disks, which is > not the case for out setup. > To give a ballpark figure: Over three days, 30 GiB were written. And the > data was not written at once, but continuously over the whole time. > > > Just to be sure, I have checked the logfiles of one of the other clusters > right now, which are sitting in the same building, in the same rack, even > on the same switch, running the same jobs, but with glusterfs 3.4.2 and I > can see no disconnects in the logfiles. So I can definitely rule out our > infrastructure as problem. > > Regards, > Micha > > > > Am 07.12.2016 um 18:08 schrieb Mohammed Rafi K C: > > Hi Micha, > > This is great. I will provide you one debug build which has two fixes > which I possible suspect for a frequent disconnect issue, though I don't > have much data to validate my theory. So I will take one more day to dig in > to that. > > Thanks for your support, and opensource++ > > Regards > > Rafi KC > On 12/07/2016 05:02 AM, Micha Ober wrote: > > Hi, > > thank you for your answer and even more for the question! > Until now, I was using FUSE. Today I changed all mounts to NFS using the > same 3.7.17 version. > > But: The problem is still the same. Now, the NFS logfile contains lines > like these: > > [2016-12-06 15:12:29.006325] C > [rpc-clnt-ping.c:165:rpc_clnt_ping_timer_expired] > 0-gv0-client-7: server X.X.18.62:49153 has not responded in the last 42 > seconds, disconnecting. > > Interestingly enough, the IP address X.X.18.62 is the same machine! As I > wrote earlier, each node serves both as a server and a client, as each node > contributes bricks to the volume. Every server is connecting to itself via > its hostname. For example, the fstab on the node "giant2" looks like: > > #giant2:/gv0 /shared_data glusterfs defaults,noauto 0 0 > #giant2:/gv2 /shared_slurm glusterfs defaults,noauto 0 0 > > giant2:/gv0 /shared_data nfs defaults,_netdev,vers=3 > 0 0 > giant2:/gv2 /shared_slurm nfs defaults,_netdev,vers=3 > 0 0 > > So I understand the disconnects even less. > > I don't know if it's possible to create a dummy cluster which exposes the > same behaviour, because the disconnects only happen when there are compute > jobs running on those nodes - and they are GPU compute jobs, so that's > something which cannot be easily emulated in a VM. > > As we have more clusters (which are running fine with an ancient 3.4 > version :-)) and we are currently not dependent on this particular cluster > (which may stay like this for this month, I think) I should be able to > deploy the debug build on the "real" cluster, if you can provide a debug > build. > > Regards and thanks, > Micha > > > > Am 06.12.2016 um 08:15 schrieb Mohammed Rafi K C: > > > > On 12/03/2016 12:56 AM, Micha Ober wrote: > > ** Update: ** I have downgraded from 3.8.6 to 3.7.17 now, but the problem > still exists. > > > Client log: http://paste.ubuntu.com/23569065/ > Brick log: http://paste.ubuntu.com/23569067/ > > Please note that each server has two bricks. > Whereas, according to the logs, one brick loses the connection to all > other hosts: > > [2016-12-02 18:38:53.703301] W [socket.c:596:__socket_rwv] 0-tcp.gv0-server: > writev on X.X.X.219:49121 failed (Broken pipe) > [2016-12-02 18:38:53.703381] W [socket.c:596:__socket_rwv] 0-tcp.gv0-server: > writev on X.X.X.62:49118 failed (Broken pipe) > [2016-12-02 18:38:53.703380] W [socket.c:596:__socket_rwv] 0-tcp.gv0-server: > writev on X.X.X.107:49121 failed (Broken pipe) > [2016-12-02 18:38:53.703424] W [socket.c:596:__socket_rwv] 0-tcp.gv0-server: > writev on X.X.X.206:49120 failed (Broken pipe) > [2016-12-02 18:38:53.703359] W [socket.c:596:__socket_rwv] 0-tcp.gv0-server: > writev on X.X.X.58:49121 failed (Broken pipe) > > The SECOND brick on the SAME host is NOT affected, i.e. no disconnects! > As I said, the network connection is fine and the disks are idle. > The CPU always has 2 free cores. > > It looks like I have to downgrade to 3.4 now in order for the disconnects to > stop. > > > Hi Micha, > > Thanks for the update and sorry for what happened with gluster higher > versions. I can understand the need for downgrade as it is a production > setup. > > Can you tell me the clients used here ? whether it is a > fuse,nfs,nfs-ganesha, smb or libgfapi ? > > Since I'm not able to reproduce the issue (I have been trying from last > 3days) and the logs are not much helpful here (we don't have much logs in > socket layer), Could you please create a dummy cluster and try to reproduce > the issue? If then we can play with that volume and I could provide some > debug build which we can use for further debugging? > > If you don't have bandwidth for this, please leave it ;). > > Regards > Rafi KC > > - Micha > > > Am 30.11.2016 um 06:57 schrieb Mohammed Rafi K C: > > Hi Micha, > > I have changed the thread and subject so that your original thread remain > same for your query. Let's try to fix the problem what you observed with > 3.8.4, So I have started a new thread to discuss the frequent disconnect > problem. > > *If any one else has experienced the same problem, please respond to the > mail.* > > It would be very helpful if you could give us some more logs from clients > and bricks. Also any reproducible steps will surely help to chase the > problem further. > > Regards > > Rafi KC > On 11/30/2016 04:44 AM, Micha Ober wrote: > > I had opened another thread on this mailing list (Subject: "After upgrade > from 3.4.2 to 3.8.5 - High CPU usage resulting in disconnects and > split-brain"). > > The title may be a bit misleading now, as I am no longer observing high > CPU usage after upgrading to 3.8.6, but the disconnects are still happening > and the number of files in split-brain is growing. > > Setup: 6 compute nodes, each serving as a glusterfs server and client, > Ubuntu 14.04, two bricks per node, distribute-replicate > > I have two gluster volumes set up (one for scratch data, one for the slurm > scheduler). Only the scratch data volume shows critical errors "[...] has > not responded in the last 42 seconds, disconnecting.". So I can rule out > network problems, the gigabit link between the nodes is not saturated at > all. The disks are almost idle (<10%). > > I have glusterfs 3.4.2 on Ubuntu 12.04 on a another compute cluster, > running fine since it was deployed. > I had glusterfs 3.4.2 on Ubuntu 14.04 on this cluster, running fine for > almost a year. > > After upgrading to 3.8.5, the problems (as described) started. I would > like to use some of the new features of the newer versions (like bitrot), > but the users can't run their compute jobs right now because the result > files are garbled. > > There also seems to be a bug report with a smiliar problem: (but no > progress) > https://bugzilla.redhat.com/show_bug.cgi?id=1370683 > > For me, ALL servers are affected (not isolated to one or two servers) > > I also see messages like "INFO: task gpu_graphene_bv:4476 blocked for > more than 120 seconds." in the syslog. > > For completeness (gv0 is the scratch volume, gv2 the slurm volume): > > [root@giant2: ~]# gluster v info > > Volume Name: gv0 > Type: Distributed-Replicate > Volume ID: 993ec7c9-e4bc-44d0-b7c4-2d977e622e86 > Status: Started > Snapshot Count: 0 > Number of Bricks: 6 x 2 = 12 > Transport-type: tcp > Bricks: > Brick1: giant1:/gluster/sdc/gv0 > Brick2: giant2:/gluster/sdc/gv0 > Brick3: giant3:/gluster/sdc/gv0 > Brick4: giant4:/gluster/sdc/gv0 > Brick5: giant5:/gluster/sdc/gv0 > Brick6: giant6:/gluster/sdc/gv0 > Brick7: giant1:/gluster/sdd/gv0 > Brick8: giant2:/gluster/sdd/gv0 > Brick9: giant3:/gluster/sdd/gv0 > Brick10: giant4:/gluster/sdd/gv0 > Brick11: giant5:/gluster/sdd/gv0 > Brick12: giant6:/gluster/sdd/gv0 > Options Reconfigured: > auth.allow: X.X.X.*,127.0.0.1 > nfs.disable: on > > Volume Name: gv2 > Type: Replicate > Volume ID: 30c78928-5f2c-4671-becc-8deaee1a7a8d > Status: Started > Snapshot Count: 0 > Number of Bricks: 1 x 2 = 2 > Transport-type: tcp > Bricks: > Brick1: giant1:/gluster/sdd/gv2 > Brick2: giant2:/gluster/sdd/gv2 > Options Reconfigured: > auth.allow: X.X.X.*,127.0.0.1 > cluster.granular-entry-heal: on > cluster.locking-scheme: granular > nfs.disable: on > > > 2016-11-30 0:10 GMT+01:00 Micha Ober <mich...@gmail.com>: > >> There also seems to be a bug report with a smiliar problem: (but no >> progress) >> https://bugzilla.redhat.com/show_bug.cgi?id=1370683 >> >> For me, ALL servers are affected (not isolated to one or two servers) >> >> I also see messages like "INFO: task gpu_graphene_bv:4476 blocked for >> more than 120 seconds." in the syslog. >> >> For completeness (gv0 is the scratch volume, gv2 the slurm volume): >> >> [root@giant2: ~]# gluster v info >> >> Volume Name: gv0 >> Type: Distributed-Replicate >> Volume ID: 993ec7c9-e4bc-44d0-b7c4-2d977e622e86 >> Status: Started >> Snapshot Count: 0 >> Number of Bricks: 6 x 2 = 12 >> Transport-type: tcp >> Bricks: >> Brick1: giant1:/gluster/sdc/gv0 >> Brick2: giant2:/gluster/sdc/gv0 >> Brick3: giant3:/gluster/sdc/gv0 >> Brick4: giant4:/gluster/sdc/gv0 >> Brick5: giant5:/gluster/sdc/gv0 >> Brick6: giant6:/gluster/sdc/gv0 >> Brick7: giant1:/gluster/sdd/gv0 >> Brick8: giant2:/gluster/sdd/gv0 >> Brick9: giant3:/gluster/sdd/gv0 >> Brick10: giant4:/gluster/sdd/gv0 >> Brick11: giant5:/gluster/sdd/gv0 >> Brick12: giant6:/gluster/sdd/gv0 >> Options Reconfigured: >> auth.allow: X.X.X.*,127.0.0.1 >> nfs.disable: on >> >> Volume Name: gv2 >> Type: Replicate >> Volume ID: 30c78928-5f2c-4671-becc-8deaee1a7a8d >> Status: Started >> Snapshot Count: 0 >> Number of Bricks: 1 x 2 = 2 >> Transport-type: tcp >> Bricks: >> Brick1: giant1:/gluster/sdd/gv2 >> Brick2: giant2:/gluster/sdd/gv2 >> Options Reconfigured: >> auth.allow: X.X.X.*,127.0.0.1 >> cluster.granular-entry-heal: on >> cluster.locking-scheme: granular >> nfs.disable: on >> >> >> 2016-11-29 19:21 GMT+01:00 Micha Ober <mich...@gmail.com>: >> >>> I had opened another thread on this mailing list (Subject: "After >>> upgrade from 3.4.2 to 3.8.5 - High CPU usage resulting in disconnects and >>> split-brain"). >>> >>> The title may be a bit misleading now, as I am no longer observing high >>> CPU usage after upgrading to 3.8.6, but the disconnects are still happening >>> and the number of files in split-brain is growing. >>> >>> Setup: 6 compute nodes, each serving as a glusterfs server and client, >>> Ubuntu 14.04, two bricks per node, distribute-replicate >>> >>> I have two gluster volumes set up (one for scratch data, one for the >>> slurm scheduler). Only the scratch data volume shows critical errors "[...] >>> has not responded in the last 42 seconds, disconnecting.". So I can rule >>> out network problems, the gigabit link between the nodes is not saturated >>> at all. The disks are almost idle (<10%). >>> >>> I have glusterfs 3.4.2 on Ubuntu 12.04 on a another compute cluster, >>> running fine since it was deployed. >>> I had glusterfs 3.4.2 on Ubuntu 14.04 on this cluster, running fine for >>> almost a year. >>> >>> After upgrading to 3.8.5, the problems (as described) started. I would >>> like to use some of the new features of the newer versions (like bitrot), >>> but the users can't run their compute jobs right now because the result >>> files are garbled. >>> >>> 2016-11-29 18:53 GMT+01:00 Atin Mukherjee <amukh...@redhat.com>: >>> >>>> Would you be able to share what is not working for you in 3.8.x >>>> (mention the exact version). 3.4 is quite old and falling back to an >>>> unsupported version doesn't look a feasible option. >>>> >>>> On Tue, 29 Nov 2016 at 17:01, Micha Ober <mich...@gmail.com> wrote: >>>> >>>>> Hi, >>>>> >>>>> I was using gluster 3.4 and upgraded to 3.8, but that version showed >>>>> to be unusable for me. I now need to downgrade. >>>>> >>>>> I'm running Ubuntu 14.04. As upgrades of the op version >>>>> are irreversible, I guess I have to delete all gluster volumes and >>>>> re-create them with the downgraded version. >>>>> >>>>> 0. Backup data >>>>> 1. Unmount all gluster volumes >>>>> 2. apt-get purge glusterfs-server glusterfs-client >>>>> 3. Remove PPA for 3.8 >>>>> 4. Add PPA for older version >>>>> 5. apt-get install glusterfs-server glusterfs-client >>>>> 6. Create volumes >>>>> >>>>> Is "purge" enough to delete all configuration files of the currently >>>>> installed version or do I need to manually clear some residues before >>>>> installing an older version? >>>>> >>>>> Thanks. >>>>> _______________________________________________ >>>>> Gluster-users mailing list >>>>> Gluster-users@gluster.org >>>>> http://www.gluster.org/mailman/listinfo/gluster-users >>>> >>>> -- >>>> - Atin (atinm) >>>> >>> >>> >> > > > _______________________________________________ > Gluster-users mailing > listGluster-users@gluster.orghttp://www.gluster.org/mailman/listinfo/gluster-users > > > > > > > > -- ~ Atin (atinm)
_______________________________________________ Gluster-users mailing list Gluster-users@gluster.org http://www.gluster.org/mailman/listinfo/gluster-users