Re: [Gluster-users] HA NFS
Hi, I've been doing some testing with ucarp and unfs, but without any luck. The VIP failover works flawlessly and clients only lose 1 ICMP echo reply while pinging the VIP during failover. However, file transfers are interrupted and never finish and the processes are hanging. Could anybody give me some more details on how to set this up? Best regards, Koen Calliauw 2010/7/14 Tejas N. Bhise > Hi Koen, > > We are infact looking at integrating this into the platform. > > Regards, > Tejas. > - Original Message - > From: "Layer7 Consultancy" > To: "Gluster General Discussion List" > Sent: Wednesday, July 14, 2010 1:06:38 PM > Subject: Re: [Gluster-users] HA NFS > > Thanks Vikas, > > I think I have enough information now to set up a proof of concept. > Perhaps this (ucarp install by user) is something to be mentioned on > the website? Even better would be integration into Gluster Platform > ;-) > > Best regards, > Koen > > 2010/7/14 Shehjar Tikoo : > > Layer7 Consultancy wrote: > >> > >> Hello Vikas, > >> > >> How do the NFS clients react to such a failover? Will the I/O just > >> temporarily stall and then proceed once connectivity is resumed or > >> will running transactions be aborted? > >> > > > > There may be some delay as the IP migrates but overall, the NFS client's > > retransmission behaviour will be able continue the transactions with no > > interruptions. > > > >> Best regards, > >> Koen > >> > >> 2010/7/13 Vikas Gorur : > >> Your understanding is correct. Whether using the native NFS > >> translator or re-export through unfsd, NFS clients only connect to the > >> management IP. If you want failover for the NFS server, you'd setup a > >> virtual IP using ucarp (http://www.ucarp.org) and the clients would > >> only use this virtual IP. > >>> > >>> -- > >>> Vikas Gorur > >>> Engineer - Gluster, Inc. > >>> -- > >> > >> ___ > >> Gluster-users mailing list > >> Gluster-users@gluster.org > >> http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > > > > ___ > > Gluster-users mailing list > > Gluster-users@gluster.org > > http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > > > ___ > Gluster-users mailing list > Gluster-users@gluster.org > http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > ___ > Gluster-users mailing list > Gluster-users@gluster.org > http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] HA NFS
Hi Koen, We are infact looking at integrating this into the platform. Regards, Tejas. - Original Message - From: "Layer7 Consultancy" To: "Gluster General Discussion List" Sent: Wednesday, July 14, 2010 1:06:38 PM Subject: Re: [Gluster-users] HA NFS Thanks Vikas, I think I have enough information now to set up a proof of concept. Perhaps this (ucarp install by user) is something to be mentioned on the website? Even better would be integration into Gluster Platform ;-) Best regards, Koen 2010/7/14 Shehjar Tikoo : > Layer7 Consultancy wrote: >> >> Hello Vikas, >> >> How do the NFS clients react to such a failover? Will the I/O just >> temporarily stall and then proceed once connectivity is resumed or >> will running transactions be aborted? >> > > There may be some delay as the IP migrates but overall, the NFS client's > retransmission behaviour will be able continue the transactions with no > interruptions. > >> Best regards, >> Koen >> >> 2010/7/13 Vikas Gorur : >> Your understanding is correct. Whether using the native NFS >> translator or re-export through unfsd, NFS clients only connect to the >> management IP. If you want failover for the NFS server, you'd setup a >> virtual IP using ucarp (http://www.ucarp.org) and the clients would >> only use this virtual IP. >>> >>> -- >>> Vikas Gorur >>> Engineer - Gluster, Inc. >>> -- >> >> ___ >> Gluster-users mailing list >> Gluster-users@gluster.org >> http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > > ___ > Gluster-users mailing list > Gluster-users@gluster.org > http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] HA NFS
Thanks Vikas, I think I have enough information now to set up a proof of concept. Perhaps this (ucarp install by user) is something to be mentioned on the website? Even better would be integration into Gluster Platform ;-) Best regards, Koen 2010/7/14 Shehjar Tikoo : > Layer7 Consultancy wrote: >> >> Hello Vikas, >> >> How do the NFS clients react to such a failover? Will the I/O just >> temporarily stall and then proceed once connectivity is resumed or >> will running transactions be aborted? >> > > There may be some delay as the IP migrates but overall, the NFS client's > retransmission behaviour will be able continue the transactions with no > interruptions. > >> Best regards, >> Koen >> >> 2010/7/13 Vikas Gorur : >> Your understanding is correct. Whether using the native NFS >> translator or re-export through unfsd, NFS clients only connect to the >> management IP. If you want failover for the NFS server, you'd setup a >> virtual IP using ucarp (http://www.ucarp.org) and the clients would >> only use this virtual IP. >>> >>> -- >>> Vikas Gorur >>> Engineer - Gluster, Inc. >>> -- >> >> ___ >> Gluster-users mailing list >> Gluster-users@gluster.org >> http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > > ___ > Gluster-users mailing list > Gluster-users@gluster.org > http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] HA NFS
Layer7 Consultancy wrote: Hello Vikas, How do the NFS clients react to such a failover? Will the I/O just temporarily stall and then proceed once connectivity is resumed or will running transactions be aborted? There may be some delay as the IP migrates but overall, the NFS client's retransmission behaviour will be able continue the transactions with no interruptions. Best regards, Koen 2010/7/13 Vikas Gorur : Your understanding is correct. Whether using the native NFS translator or re-export through unfsd, NFS clients only connect to the management IP. If you want failover for the NFS server, you'd setup a virtual IP using ucarp (http://www.ucarp.org) and the clients would only use this virtual IP. -- Vikas Gorur Engineer - Gluster, Inc. -- ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] HA NFS
Hello Vikas, How do the NFS clients react to such a failover? Will the I/O just temporarily stall and then proceed once connectivity is resumed or will running transactions be aborted? Best regards, Koen 2010/7/13 Vikas Gorur : Your understanding is correct. Whether using the native NFS translator or re-export through unfsd, NFS clients only connect to the management IP. If you want failover for the NFS server, you'd setup a virtual IP using ucarp (http://www.ucarp.org) and the clients would only use this virtual IP. > > -- > Vikas Gorur > Engineer - Gluster, Inc. > -- ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] HA NFS
On Jul 13, 2010, at 7:58 AM, Layer7 Consultancy wrote: > Hi Chris, > > Thanks for your answer, this was also my understanding on how things > work when using the FUSE client (all storage nodes are described in > the glusterfsd.vol file and thus the client has connection info to > them). > > However, when using NFS the documentation states that one should > connect to the 'management IP' and this also seems to be the only > connection information that the client has. > If this management IP is gone due to the server going down, there is > no way the client can know there are multiple other servers who are > also serving the same content, so unless this virtual IP is taken over > by another storage node, the client wouldn't know where to route the > request to. > > Can anyone confirm this? Your understanding is correct. Whether using the native NFS translator or re-export through unfsd, NFS clients only connect to the management IP. If you want failover for the NFS server, you'd setup a virtual IP using ucarp (http://www.ucarp.org) and the clients would only use this virtual IP. -- Vikas Gorur Engineer - Gluster, Inc. -- ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] HA NFS
Ah, I had no idea! I look forward to the correct answer because I am interested in the NFS translator as well. - "Layer7 Consultancy" wrote: > Hi Chris, > > Thanks for your answer, this was also my understanding on how things > work when using the FUSE client (all storage nodes are described in > the glusterfsd.vol file and thus the client has connection info to > them). > > However, when using NFS the documentation states that one should > connect to the 'management IP' and this also seems to be the only > connection information that the client has. > If this management IP is gone due to the server going down, there is > no way the client can know there are multiple other servers who are > also serving the same content, so unless this virtual IP is taken > over > by another storage node, the client wouldn't know where to route the > request to. > > Can anyone confirm this? > > Cheers, > Koen > > > 2010/7/13 Christopher Hawkins : > > I can offer a little general information. My understanding is this: > > > > It is not like failover with a virtual IP. Instead, the gluster > clients connect to all storage servers at the same time. If one of > them becomes unavailable, the client can still reach the remaining > one(s). Locks are preserved for all remaining nodes. Writes are marked > (in the metadata) as having been completed on the remaining nodes, and > NOT completed on whatever nodes is down. On access, the file will be > healed if the downed node has returned. Or you can force healing of > all files when the node comes back, simply by accessing all files with > a 'find' command. See seal healing in the wiki for more information on > this. > > > > I am not familiar with OpenQRM so I don't know if or how that would > have be tweaked for integration. > > > > Chris > > > > - "Layer7 Consultancy" wrote: > > > >> Hi all, > >> > >> I am considering the built-in NFS functionality of Gluster to build > a > >> virtual server environment. The idea is to have 4 or 5 hosts (KVM > or > >> Xen) that all contain around 300GB of 15K rpm SAS storage in a > RAID5 > >> array. On each of the host servers I would install a VM with the > >> Gluster Platform and expose all of this storage through NFS to my > >> OpenQRM installation, which would then host all the other VM's on > the > >> same servers. > >> An alternative idea is to have the storage boxes separate from the > VM > >> hosts, but the basic idea stays the same I think. > >> > >> Now from what I understand, the NFS storage that is exposed to the > >> clients is approached through the management IP of the first > Gluster > >> Platform server. My biggest question is what exactly happens when > the > >> first storage node goes down. Does the platform offer some kind of > >> VRRP setup that fails over the IP to one of the other nodes? Is > the > >> lock information preserved and how does this all work internally? > >> > >> Since I would be using KVM or Xen, it would in theory be possible > to > >> build the FUSE client on the host servers, though I am still in > doubt > >> on how OpenQRM will handle this. When choosing local storage, > OpenQRM > >> expects raw disks (I think) and creates LVM groups on these disks > in > >> order to allow snapshotting and backups. OpenQRM would also not > know > >> this is shared storage. > >> > >> Does anyone have some insight on a setup like this? > >> > >> Best regards, > >> Koen > >> ___ > >> Gluster-users mailing list > >> Gluster-users@gluster.org > >> http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > > ___ > > Gluster-users mailing list > > Gluster-users@gluster.org > > http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > > > ___ > Gluster-users mailing list > Gluster-users@gluster.org > http://gluster.org/cgi-bin/mailman/listinfo/gluster-users ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] HA NFS
Hi Chris, Thanks for your answer, this was also my understanding on how things work when using the FUSE client (all storage nodes are described in the glusterfsd.vol file and thus the client has connection info to them). However, when using NFS the documentation states that one should connect to the 'management IP' and this also seems to be the only connection information that the client has. If this management IP is gone due to the server going down, there is no way the client can know there are multiple other servers who are also serving the same content, so unless this virtual IP is taken over by another storage node, the client wouldn't know where to route the request to. Can anyone confirm this? Cheers, Koen 2010/7/13 Christopher Hawkins : > I can offer a little general information. My understanding is this: > > It is not like failover with a virtual IP. Instead, the gluster clients > connect to all storage servers at the same time. If one of them becomes > unavailable, the client can still reach the remaining one(s). Locks are > preserved for all remaining nodes. Writes are marked (in the metadata) as > having been completed on the remaining nodes, and NOT completed on whatever > nodes is down. On access, the file will be healed if the downed node has > returned. Or you can force healing of all files when the node comes back, > simply by accessing all files with a 'find' command. See seal healing in the > wiki for more information on this. > > I am not familiar with OpenQRM so I don't know if or how that would have be > tweaked for integration. > > Chris > > - "Layer7 Consultancy" wrote: > >> Hi all, >> >> I am considering the built-in NFS functionality of Gluster to build a >> virtual server environment. The idea is to have 4 or 5 hosts (KVM or >> Xen) that all contain around 300GB of 15K rpm SAS storage in a RAID5 >> array. On each of the host servers I would install a VM with the >> Gluster Platform and expose all of this storage through NFS to my >> OpenQRM installation, which would then host all the other VM's on the >> same servers. >> An alternative idea is to have the storage boxes separate from the VM >> hosts, but the basic idea stays the same I think. >> >> Now from what I understand, the NFS storage that is exposed to the >> clients is approached through the management IP of the first Gluster >> Platform server. My biggest question is what exactly happens when the >> first storage node goes down. Does the platform offer some kind of >> VRRP setup that fails over the IP to one of the other nodes? Is the >> lock information preserved and how does this all work internally? >> >> Since I would be using KVM or Xen, it would in theory be possible to >> build the FUSE client on the host servers, though I am still in doubt >> on how OpenQRM will handle this. When choosing local storage, OpenQRM >> expects raw disks (I think) and creates LVM groups on these disks in >> order to allow snapshotting and backups. OpenQRM would also not know >> this is shared storage. >> >> Does anyone have some insight on a setup like this? >> >> Best regards, >> Koen >> ___ >> Gluster-users mailing list >> Gluster-users@gluster.org >> http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > ___ > Gluster-users mailing list > Gluster-users@gluster.org > http://gluster.org/cgi-bin/mailman/listinfo/gluster-users > ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
Re: [Gluster-users] HA NFS
I can offer a little general information. My understanding is this: It is not like failover with a virtual IP. Instead, the gluster clients connect to all storage servers at the same time. If one of them becomes unavailable, the client can still reach the remaining one(s). Locks are preserved for all remaining nodes. Writes are marked (in the metadata) as having been completed on the remaining nodes, and NOT completed on whatever nodes is down. On access, the file will be healed if the downed node has returned. Or you can force healing of all files when the node comes back, simply by accessing all files with a 'find' command. See seal healing in the wiki for more information on this. I am not familiar with OpenQRM so I don't know if or how that would have be tweaked for integration. Chris - "Layer7 Consultancy" wrote: > Hi all, > > I am considering the built-in NFS functionality of Gluster to build a > virtual server environment. The idea is to have 4 or 5 hosts (KVM or > Xen) that all contain around 300GB of 15K rpm SAS storage in a RAID5 > array. On each of the host servers I would install a VM with the > Gluster Platform and expose all of this storage through NFS to my > OpenQRM installation, which would then host all the other VM's on the > same servers. > An alternative idea is to have the storage boxes separate from the VM > hosts, but the basic idea stays the same I think. > > Now from what I understand, the NFS storage that is exposed to the > clients is approached through the management IP of the first Gluster > Platform server. My biggest question is what exactly happens when the > first storage node goes down. Does the platform offer some kind of > VRRP setup that fails over the IP to one of the other nodes? Is the > lock information preserved and how does this all work internally? > > Since I would be using KVM or Xen, it would in theory be possible to > build the FUSE client on the host servers, though I am still in doubt > on how OpenQRM will handle this. When choosing local storage, OpenQRM > expects raw disks (I think) and creates LVM groups on these disks in > order to allow snapshotting and backups. OpenQRM would also not know > this is shared storage. > > Does anyone have some insight on a setup like this? > > Best regards, > Koen > ___ > Gluster-users mailing list > Gluster-users@gluster.org > http://gluster.org/cgi-bin/mailman/listinfo/gluster-users ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
[Gluster-users] HA NFS
Hi all, I am considering the built-in NFS functionality of Gluster to build a virtual server environment. The idea is to have 4 or 5 hosts (KVM or Xen) that all contain around 300GB of 15K rpm SAS storage in a RAID5 array. On each of the host servers I would install a VM with the Gluster Platform and expose all of this storage through NFS to my OpenQRM installation, which would then host all the other VM's on the same servers. An alternative idea is to have the storage boxes separate from the VM hosts, but the basic idea stays the same I think. Now from what I understand, the NFS storage that is exposed to the clients is approached through the management IP of the first Gluster Platform server. My biggest question is what exactly happens when the first storage node goes down. Does the platform offer some kind of VRRP setup that fails over the IP to one of the other nodes? Is the lock information preserved and how does this all work internally? Since I would be using KVM or Xen, it would in theory be possible to build the FUSE client on the host servers, though I am still in doubt on how OpenQRM will handle this. When choosing local storage, OpenQRM expects raw disks (I think) and creates LVM groups on these disks in order to allow snapshotting and backups. OpenQRM would also not know this is shared storage. Does anyone have some insight on a setup like this? Best regards, Koen ___ Gluster-users mailing list Gluster-users@gluster.org http://gluster.org/cgi-bin/mailman/listinfo/gluster-users