Hi, We''re setting up fairly large Lustre 2.1.2 filesystems, each with 18 nodes and 159 resources all in one Corosync/Pacemaker cluster as suggested by our vendor. We''re getting mixed messages on how large of a Corosync/Pacemaker cluster will work well between our vendor an others. 1. Are there Lustre Corosync/Pacemaker clusters out there of this size or larger? 2. If so, what tuning needed to be done to get it to work well? 3. Should we be looking more seriously into splitting this Corosync/Pacemaker cluster into pairs or sets of 4 nodes? Right now, our current configuration takes a long time to start/stop all resources (~30-45 mins), and failing back OSTs puts a heavy load on the cib process on every node in the cluster. Under heavy IO load, the many of the nodes will show as "unclean/offline" and many OST resources will show as inactive in crm status, despite the fact that every single MDT and OST is still mounted in the appropriate place. We are running 2 corosync rings, each on a private 1 GbE network. We have a bonded 10 GbE network for the LNET. Thanks, Shawn -------------- next part -------------- An HTML attachment was scrubbed... URL: http://lists.lustre.org/pipermail/lustre-discuss/attachments/20121019/46326bb4/attachment.html
FWIW, we are running HA Lustre using corosync/pacemaker. We broke our OSSs and MDSs out into individual HA *pairs*. Thought about other configurations but it was our first step into corosync/pacemaker so we decided to keep it as simple as possible. Seems to work well. I''m not sure I would attempt what you are doing though it may be perfectly fine. When HA is a requirement, it probably makes sense to avoid pushing the limits of what works. Doesn''t really help you much other than to provide a data point with regard to what other sites are doing. Good luck and report back. Charlie Taylor UF HPC Center On Oct 19, 2012, at 12:52 PM, Hall, Shawn wrote:> Hi, > > We?re setting up fairly large Lustre 2.1.2 filesystems, each with 18 nodes and 159 resources all in one Corosync/Pacemaker cluster as suggested by our vendor. We?re getting mixed messages on how large of a Corosync/Pacemaker cluster will work well between our vendor an others. > > 1. Are there Lustre Corosync/Pacemaker clusters out there of this size or larger? > 2. If so, what tuning needed to be done to get it to work well? > 3. Should we be looking more seriously into splitting this Corosync/Pacemaker cluster into pairs or sets of 4 nodes? > > Right now, our current configuration takes a long time to start/stop all resources (~30-45 mins), and failing back OSTs puts a heavy load on the cib process on every node in the cluster. Under heavy IO load, the many of the nodes will show as ?unclean/offline? and many OST resources will show as inactive in crm status, despite the fact that every single MDT and OST is still mounted in the appropriate place. We are running 2 corosync rings, each on a private 1 GbE network. We have a bonded 10 GbE network for the LNET. > > Thanks, > Shawn > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss at lists.lustre.org > http://lists.lustre.org/mailman/listinfo/lustre-discuss
Shawn, In my opinion you shouldn''t be running corosync on any more than two machines. They should be configured in self contained pairs (mds pair, oss pairs). Anything beyond that would be chaos to manage, even if it worked. Don''t forget the stonith portion. Not every block storage implementation respects mmp protection. --Jeff On 10/19/12 9:52 AM, Hall, Shawn wrote:> > Hi, > > We?re setting up fairly large Lustre 2.1.2 filesystems, each with 18 > nodes and 159 resources all in one Corosync/Pacemaker cluster as > suggested by our vendor. We?re getting mixed messages on how large of > a Corosync/Pacemaker cluster will work well between our vendor an others. > > 1.Are there Lustre Corosync/Pacemaker clusters out there of this size > or larger? > > 2.If so, what tuning needed to be done to get it to work well? > > 3.Should we be looking more seriously into splitting this > Corosync/Pacemaker cluster into pairs or sets of 4 nodes? > > Right now, our current configuration takes a long time to start/stop > all resources (~30-45 mins), and failing back OSTs puts a heavy load > on the cib process on every node in the cluster. Under heavy IO load, > the many of the nodes will show as ?unclean/offline? and many OST > resources will show as inactive in crm status, despite the fact that > every single MDT and OST is still mounted in the appropriate place. We > are running 2 corosync rings, each on a private 1 GbE network. We have > a bonded 10 GbE network for the LNET. > > Thanks, > > Shawn > > > > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss at lists.lustre.org > http://lists.lustre.org/mailman/listinfo/lustre-discuss-- ------------------------------ Jeff Johnson Co-Founder Aeon Computing jeff.johnson at aeoncomputing.com www.aeoncomputing.com t: 858-412-3810 x101 f: 858-412-3845 m: 619-204-9061 /* New Address */ 4170 Morena Boulevard, Suite D - San Diego, CA 92117
Thanks for the replies. We''ve worked on the HA and have it to a satisfactory point where we can put it into production. We broke it into a MDS pair and 4 groups of 4 OSS nodes. From our perspective, it''s actually easier to manage groups of 4 than groups of 2, since it''s half as many configurations to keep track of. After splitting the cluster into 5 pieces it has become much more responsive and stable. It''s more difficult to manage than one large cluster, but the stability is obviously worth it. We''ve been performing heavy load testing and have not been able to "break" the cluster. We did a few more things to get to this point: - Lowered the nice value of the corosync process to make it more responsive under load and prevent a node from getting kicked out due to unresponsiveness. - Increased vm.min_free_kbytes to give TCP/IP w/ jumbo frames room to move around. Without this certain nodes would have low memory issues related to networking and would get stonithed due to unresponsiveness. Thanks, Shawn -----Original Message----- From: Charles Taylor [mailto:taylor at hpc.ufl.edu] Sent: Wednesday, October 24, 2012 3:33 PM To: Hall, Shawn Cc: lustre-discuss at lists.lustre.org Subject: Re: [Lustre-discuss] Large Corosync/Pacemaker clusters FWIW, we are running HA Lustre using corosync/pacemaker. We broke our OSSs and MDSs out into individual HA *pairs*. Thought about other configurations but it was our first step into corosync/pacemaker so we decided to keep it as simple as possible. Seems to work well. I''m not sure I would attempt what you are doing though it may be perfectly fine. When HA is a requirement, it probably makes sense to avoid pushing the limits of what works. Doesn''t really help you much other than to provide a data point with regard to what other sites are doing. Good luck and report back. Charlie Taylor UF HPC Center On Oct 19, 2012, at 12:52 PM, Hall, Shawn wrote:> Hi, > > We''re setting up fairly large Lustre 2.1.2 filesystems, each with 18nodes and 159 resources all in one Corosync/Pacemaker cluster as suggested by our vendor. We''re getting mixed messages on how large of a Corosync/Pacemaker cluster will work well between our vendor an others.> > 1. Are there Lustre Corosync/Pacemaker clusters out there ofthis size or larger?> 2. If so, what tuning needed to be done to get it to work well? > 3. Should we be looking more seriously into splitting thisCorosync/Pacemaker cluster into pairs or sets of 4 nodes?> > Right now, our current configuration takes a long time to start/stopall resources (~30-45 mins), and failing back OSTs puts a heavy load on the cib process on every node in the cluster. Under heavy IO load, the many of the nodes will show as "unclean/offline" and many OST resources will show as inactive in crm status, despite the fact that every single MDT and OST is still mounted in the appropriate place. We are running 2 corosync rings, each on a private 1 GbE network. We have a bonded 10 GbE network for the LNET.> > Thanks, > Shawn > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss at lists.lustre.org > http://lists.lustre.org/mailman/listinfo/lustre-discuss
Hi, I''m also setting up a high-available Lustre system, I configured pairs for the OSSes and MDSes, redundant Corosync rings (two separate rings: IB and Eth), and Stonith is enabled. The current configuration seems to work fine, however yesterday we experienced some problem because 4 OSSes got rebooted by Stonith. I suspect that Corosync missed a heartbeat due to a kernel/corosync hung, rather than a network problem. I will try the "renice" solution you proposed. I have been thinking that I could increase the "token" timeout value in /etc/corosync/corosync.conf , to prevent short "hiccups". Did you specify a value to this parameter or did you leave the default 1000ms value? Marco On 2012-10-31 03:43, Hall, Shawn wrote:> Thanks for the replies. We''ve worked on the HA and have it to a > satisfactory point where we can put it into production. We broke it > into a MDS pair and 4 groups of 4 OSS nodes. From our perspective, it''s > actually easier to manage groups of 4 than groups of 2, since it''s half > as many configurations to keep track of. > > After splitting the cluster into 5 pieces it has become much more > responsive and stable. It''s more difficult to manage than one large > cluster, but the stability is obviously worth it. We''ve been performing > heavy load testing and have not been able to "break" the cluster. We > did a few more things to get to this point: > > - Lowered the nice value of the corosync process to make it more > responsive under load and prevent a node from getting kicked out due to > unresponsiveness. > - Increased vm.min_free_kbytes to give TCP/IP w/ jumbo frames room to > move around. Without this certain nodes would have low memory issues > related to networking and would get stonithed due to unresponsiveness. > > Thanks, > Shawn > > -----Original Message----- > From: Charles Taylor [mailto:taylor at hpc.ufl.edu] > Sent: Wednesday, October 24, 2012 3:33 PM > To: Hall, Shawn > Cc: lustre-discuss at lists.lustre.org > Subject: Re: [Lustre-discuss] Large Corosync/Pacemaker clusters > > > FWIW, we are running HA Lustre using corosync/pacemaker. We broke our > OSSs and MDSs out into individual HA *pairs*. Thought about other > configurations but it was our first step into corosync/pacemaker so we > decided to keep it as simple as possible. Seems to work well. I''m > not sure I would attempt what you are doing though it may be perfectly > fine. When HA is a requirement, it probably makes sense to avoid > pushing the limits of what works. > > Doesn''t really help you much other than to provide a data point with > regard to what other sites are doing. > > Good luck and report back. > > Charlie Taylor > UF HPC Center > > On Oct 19, 2012, at 12:52 PM, Hall, Shawn wrote: > >> Hi, >> >> We''re setting up fairly large Lustre 2.1.2 filesystems, each with 18 > nodes and 159 resources all in one Corosync/Pacemaker cluster as > suggested by our vendor. We''re getting mixed messages on how large of a > Corosync/Pacemaker cluster will work well between our vendor an others. >> >> 1. Are there Lustre Corosync/Pacemaker clusters out there of > this size or larger? >> 2. If so, what tuning needed to be done to get it to work well? >> 3. Should we be looking more seriously into splitting this > Corosync/Pacemaker cluster into pairs or sets of 4 nodes? >> >> Right now, our current configuration takes a long time to start/stop > all resources (~30-45 mins), and failing back OSTs puts a heavy load on > the cib process on every node in the cluster. Under heavy IO load, the > many of the nodes will show as "unclean/offline" and many OST resources > will show as inactive in crm status, despite the fact that every single > MDT and OST is still mounted in the appropriate place. We are running 2 > corosync rings, each on a private 1 GbE network. We have a bonded 10 > GbE network for the LNET. >> >> Thanks, >> Shawn >> _______________________________________________ >> Lustre-discuss mailing list >> Lustre-discuss at lists.lustre.org >> http://lists.lustre.org/mailman/listinfo/lustre-discuss > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss at lists.lustre.org > http://lists.lustre.org/mailman/listinfo/lustre-discuss
Hi, Our vendor actually has several of the parameters in corosync.conf increased by default, and we have not touched them. These are: Token: 10000 Retransmits_before_loss: 25 Consensus: 12000 Join: 1000 Merge: 400 Downcheck: 2000 We also have secauth turned off, as this can consume 75% of your CPU cycles and cut bandwidth by a third, according to the corosync.conf manpage. I''m not sure if these parameters are necessary now that we have split our cluster up, but they haven''t seemed to hurt anything either. Hope this helps, Shawn -----Original Message----- From: Marco Passerini [mailto:marco.passerini at csc.fi] Sent: Tuesday, November 06, 2012 7:13 AM To: lustre-discuss at lists.lustre.org Cc: Hall, Shawn Subject: Re: [Lustre-discuss] Large Corosync/Pacemaker clusters Hi, I''m also setting up a high-available Lustre system, I configured pairs for the OSSes and MDSes, redundant Corosync rings (two separate rings: IB and Eth), and Stonith is enabled. The current configuration seems to work fine, however yesterday we experienced some problem because 4 OSSes got rebooted by Stonith. I suspect that Corosync missed a heartbeat due to a kernel/corosync hung, rather than a network problem. I will try the "renice" solution you proposed. I have been thinking that I could increase the "token" timeout value in /etc/corosync/corosync.conf , to prevent short "hiccups". Did you specify a value to this parameter or did you leave the default 1000ms value? Marco On 2012-10-31 03:43, Hall, Shawn wrote:> Thanks for the replies. We''ve worked on the HA and have it to a > satisfactory point where we can put it into production. We broke it > into a MDS pair and 4 groups of 4 OSS nodes. From our perspective, > it''s actually easier to manage groups of 4 than groups of 2, since > it''s half as many configurations to keep track of. > > After splitting the cluster into 5 pieces it has become much more > responsive and stable. It''s more difficult to manage than one large > cluster, but the stability is obviously worth it. We''ve been > performing heavy load testing and have not been able to "break" the > cluster. We did a few more things to get to this point: > > - Lowered the nice value of the corosync process to make it more > responsive under load and prevent a node from getting kicked out due > to unresponsiveness. > - Increased vm.min_free_kbytes to give TCP/IP w/ jumbo frames room to > move around. Without this certain nodes would have low memory issues > related to networking and would get stonithed due to unresponsiveness. > > Thanks, > Shawn > > -----Original Message----- > From: Charles Taylor [mailto:taylor at hpc.ufl.edu] > Sent: Wednesday, October 24, 2012 3:33 PM > To: Hall, Shawn > Cc: lustre-discuss at lists.lustre.org > Subject: Re: [Lustre-discuss] Large Corosync/Pacemaker clusters > > > FWIW, we are running HA Lustre using corosync/pacemaker. We brokeour> OSSs and MDSs out into individual HA *pairs*. Thought about other > configurations but it was our first step into corosync/pacemaker so we > decided to keep it as simple as possible. Seems to work well. I''m > not sure I would attempt what you are doing though it may be perfectly > fine. When HA is a requirement, it probably makes sense to avoid > pushing the limits of what works. > > Doesn''t really help you much other than to provide a data point with > regard to what other sites are doing. > > Good luck and report back. > > Charlie Taylor > UF HPC Center > > On Oct 19, 2012, at 12:52 PM, Hall, Shawn wrote: > >> Hi, >> >> We''re setting up fairly large Lustre 2.1.2 filesystems, each with 18 > nodes and 159 resources all in one Corosync/Pacemaker cluster as > suggested by our vendor. We''re getting mixed messages on how large of> a Corosync/Pacemaker cluster will work well between our vendor anothers.>> >> 1. Are there Lustre Corosync/Pacemaker clusters out there of > this size or larger? >> 2. If so, what tuning needed to be done to get it to work well? >> 3. Should we be looking more seriously into splitting this > Corosync/Pacemaker cluster into pairs or sets of 4 nodes? >> >> Right now, our current configuration takes a long time to start/stop > all resources (~30-45 mins), and failing back OSTs puts a heavy load > on the cib process on every node in the cluster. Under heavy IO load,> the many of the nodes will show as "unclean/offline" and many OST > resources will show as inactive in crm status, despite the fact that > every single MDT and OST is still mounted in the appropriate place. > We are running 2 corosync rings, each on a private 1 GbE network. We > have a bonded 10 GbE network for the LNET. >> >> Thanks, >> Shawn >> _______________________________________________ >> Lustre-discuss mailing list >> Lustre-discuss at lists.lustre.org >> http://lists.lustre.org/mailman/listinfo/lustre-discuss > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss at lists.lustre.org > http://lists.lustre.org/mailman/listinfo/lustre-discuss
> I will try the "renice" solution you proposed.re-niceing corosync should not be required as the process is supposed to run with RT-Priority anyway.> I have been thinking that I could increase the "token" timeout value in > /etc/corosync/corosync.conf , to prevent short "hiccups". Did you > specify a value to this parameter or did you leave the default 1000ms value?We configured the token timeout to 17 seconds: totem { [....] transport: udpu rrp_mode: passive token: 17000 } This configuration works just fine for us since months: We didn''t see a single ''false positive STONITH'' with this configuration. Regards, Adrian