Thread: [SSI] Problem Solved Thanks more talk
Brought to you by:
brucewalker,
rogertsang
From: j g. <rav...@ya...> - 2002-03-20 21:57:44
|
Thanks Brian for the suggestion. I am embarrased ! The problem was like you said, check LABEL=/ in fstab. I was using /dev/sda1 which for some reason reading the instructions, I thought you are supposed to name your Fibre Channel here, not the disk in the server/node. I am using VA linux machines with IDE hard disk in them. so I when the piano finally fell on my head I changed the label to /dev/hda2 and things have been great. I have worked up to the part of adding a node now. I want to clear up one thing. Is it necessary for each node/server to have an internal hard disk, or can a node boot off a floppy, find the Fibre Channel Shared disk and forget about having an internal disk alltogether ? In this case, when the machine gets rebooted, we just save the floppy, or make lots of copies for other machines to have handy ? Also, each node need a Fibre Channel HBA, but does the lock server need an HBA ? Thanks for helping out.. jeff __________________________________________________ Do You Yahoo!? Yahoo! Movies - coverage of the 74th Academy Awards® http://movies.yahoo.com/ |
From: Brian J. W. <Bri...@co...> - 2002-03-20 22:58:53
|
j gage wrote: > > Thanks Brian for the suggestion. > > I am embarrased ! The problem was like you said, check > LABEL=/ in fstab. Cool. I'm glad it works. > I was using /dev/sda1 which for some reason reading > the instructions, I thought you are supposed to name > your Fibre Channel here, not the disk in the > server/node. I am using VA linux machines with IDE > hard disk in them. so I when the piano finally fell on > my head I changed the label to /dev/hda2 and things > have been great. I added a sentence to that step to try to clarify things for future users. > I have worked up to the part of adding a node now. I > want to clear up one thing. Is it necessary for each > node/server to have an internal hard disk, or can a > node boot off a floppy, find the Fibre Channel Shared > disk and forget about having an internal disk > alltogether ? Most nodes can boot with the Etherboot floppy, find the current master node, and download the kernel and ramdisk they need to join the cluster and mount the shared root. Only potential master nodes require a local boot disk, because a potential master node can be the first node booted in the cluster. In general, nodes 1 and 2 should be potential master nodes with local boot devices. The rest of the nodes do not need to be potential master nodes, and can use an Etherboot floppy. Right now, all nodes need a swap device, and it must be the same device name on all nodes (e.g., /dev/hda3). I'd like to loosen that restriction in the future. Then you could have nodes with no internal disks, so long as they have enough memory to not need swap. > In this case, when the machine gets > rebooted, we just save the floppy, or make lots of > copies for other machines to have handy ? It's best to make a copy for each node, and keep that copy in the node's floppy drive. That way, the node can simply reboot and rejoin the cluster without manual intervention. > Also, each node need a Fibre Channel HBA, but does the > lock server need an HBA ? No. It's just a dumb lock server with no understanding of GFS. > Thanks for helping out.. No problem. -- Brian Watson | "Now I don't know, but I been told it's Linux Kernel Developer | hard to run with the weight of gold, Open SSI Clustering Project | Other hand I heard it said, it's Compaq Computer Corp | just as hard with the weight of lead." Los Angeles, CA | -Robert Hunter, 1970 mailto:Bri...@co... http://opensource.compaq.com/ |
From: j g. <rav...@ya...> - 2002-03-20 23:44:26
|
Excellent ! From a cost point of view, since all nodes need a swap device, they should all be made potential masters if the disk is big enough. I have 6.4 Gigabyte IDE drives in the va linux servers, so using the whole thing for swap is a waste. Your paying for the disk already, right ? Another dumb question for the group - would there be any issues using the lock server as the Virtual Server Master and Firewall as well ? So we have 3 fucntions in one to save on expensive machines and rack space. I am just very excited about this package (SSI) ! I thought I would have to build GFS and VSP separately, but this brings it together, so I am going to try an become an expert at it. Jeff --- "Brian J. Watson" <Bri...@co...> wrote: > j gage wrote: > > > > Thanks Brian for the suggestion. > > > > I am embarrased ! The problem was like you said, > check > > LABEL=/ in fstab. > > Cool. I'm glad it works. > > > I was using /dev/sda1 which for some reason > reading > > the instructions, I thought you are supposed to > name > > your Fibre Channel here, not the disk in the > > server/node. I am using VA linux machines with IDE > > hard disk in them. so I when the piano finally > fell on > > my head I changed the label to /dev/hda2 and > things > > have been great. > > I added a sentence to that step to try to clarify > things for future > users. > > > > I have worked up to the part of adding a node now. > I > > want to clear up one thing. Is it necessary for > each > > node/server to have an internal hard disk, or can > a > > node boot off a floppy, find the Fibre Channel > Shared > > disk and forget about having an internal disk > > alltogether ? > > Most nodes can boot with the Etherboot floppy, find > the current master > node, and download the kernel and ramdisk they need > to join the cluster > and mount the shared root. Only potential master > nodes require a local > boot disk, because a potential master node can be > the first node booted > in the cluster. > > In general, nodes 1 and 2 should be potential master > nodes with local > boot devices. The rest of the nodes do not need to > be potential master > nodes, and can use an Etherboot floppy. > > Right now, all nodes need a swap device, and it must > be the same device > name on all nodes (e.g., /dev/hda3). I'd like to > loosen that restriction > in the future. Then you could have nodes with no > internal disks, so long > as they have enough memory to not need swap. > > > In this case, when the machine gets > > rebooted, we just save the floppy, or make lots of > > copies for other machines to have handy ? > > It's best to make a copy for each node, and keep > that copy in the node's > floppy drive. That way, the node can simply reboot > and rejoin the > cluster without manual intervention. > > > Also, each node need a Fibre Channel HBA, but does > the > > lock server need an HBA ? > > No. It's just a dumb lock server with no > understanding of GFS. > > > Thanks for helping out.. > > No problem. > > -- > Brian Watson | "Now I don't know, but > I been told it's > Linux Kernel Developer | hard to run with the > weight of gold, > Open SSI Clustering Project | Other hand I heard it > said, it's > Compaq Computer Corp | just as hard with the > weight of lead." > Los Angeles, CA | -Robert Hunter, > 1970 > > mailto:Bri...@co... > http://opensource.compaq.com/ __________________________________________________ Do You Yahoo!? Yahoo! Movies - coverage of the 74th Academy Awards® http://movies.yahoo.com/ |
From: Brian J. W. <Bri...@co...> - 2002-03-21 02:41:13
|
j gage wrote: > >From a cost point of view, since all nodes need a swap > device, they should all be made potential masters if > the disk is big enough. A node can have a local boot device without being a potential master. I'm not sure what you save by doing this. I think each additional potential master adds some node monitoring overhead to the interconnect. The significance of this depends on the frequency you configure for node monitor probes. > I have 6.4 Gigabyte IDE drives > in the va linux servers, so using the whole thing for > swap is a waste. Your paying for the disk already, > right ? That's true, but more nodes with local boot devices means cluster_lilo takes longer to run. > Another dumb question for the group - would there be > any issues using the lock server as the Virtual Server > Master and Firewall as well ? So we have 3 fucntions > in one to save on expensive machines and rack space. Sure. Just make sure memexpd is not bogging down the system too much. Eventually, the lock server machine will be eliminated. IBM's Distributed Lock Manager will be used to distribute the lock server across the cluster. Not only will it avoid wasting a machine, but it will also eliminate a single point of failure. Kai's put some effort into integrating Linux Virtual Server with SSI clustering. It may be preferable to run the LVS master in the cluster. Kai? > I am just very excited about this package (SSI) ! I > thought I would have to build GFS and VSP separately, > but this brings it together, so I am going to try an > become an expert at it. Thanks! It'll continue to improve over time. Any contributions are appreciated, including good bug reports. -- Brian Watson | "Now I don't know, but I been told it's Linux Kernel Developer | hard to run with the weight of gold, Open SSI Clustering Project | Other hand I heard it said, it's Compaq Computer Corp | just as hard with the weight of lead." Los Angeles, CA | -Robert Hunter, 1970 mailto:Bri...@co... http://opensource.compaq.com/ |
From: Bruce W. <br...@ka...> - 2002-03-21 02:51:15
|
> Excellent ! >=20 > From a cost point of view, since all nodes need a swap > device, they should all be made potential masters if > the disk is big enough. I have 6.4 Gigabyte IDE drives > in the va linux servers, so using the whole thing for > swap is a waste. Your paying for the disk already, > right ? right. >=20 > Another dumb question for the group - would there be > any issues using the lock server as the Virtual Server > Master and Firewall as well ? So we have 3 fucntions > in one to save on expensive machines and rack space.=20 You can certainly do that. In the future, we would=20 like to integrate the LVS master (ldirectord) more with the membership code and do failover as needed. That would require the LVS master to be in the cluster. For now, the external machine with memexpd is a single-point-of-failure anyway so you might as well put the LVS master on it. The hope w.r.t. memexpd is to move to DLM, which could be run inside the cluster and wouldn't be a single point of failure. >=20 > I am just very excited about this package (SSI) ! I > thought I would have to build GFS and VSP separately, > but this brings it together, so I am going to try an > become an expert at it. >=20 wonderful to have you on board. Please keep sharing your experiences (good and bad) so we can guide the project in a useful direction and improve the experience for all. bruce >=20 > Jeff > --- "Brian J. Watson" <Bri...@co...> > wrote: > > j gage wrote: > > >=20 > > > Thanks Brian for the suggestion. > > >=20 > > > I am embarrased ! The problem was like you said, > > check > > > LABEL=3D/ in fstab. > >=20 > > Cool. I'm glad it works. > >=20 > > > I was using /dev/sda1 which for some reason > > reading > > > the instructions, I thought you are supposed to > > name > > > your Fibre Channel here, not the disk in the > > > server/node. I am using VA linux machines with IDE > > > hard disk in them. so I when the piano finally > > fell on > > > my head I changed the label to /dev/hda2 and > > things > > > have been great. > >=20 > > I added a sentence to that step to try to clarify > > things for future > > users. > >=20 > >=20 > > > I have worked up to the part of adding a node now. > > I > > > want to clear up one thing. Is it necessary for > > each > > > node/server to have an internal hard disk, or can > > a > > > node boot off a floppy, find the Fibre Channel > > Shared > > > disk and forget about having an internal disk > > > alltogether ? > >=20 > > Most nodes can boot with the Etherboot floppy, find > > the current master > > node, and download the kernel and ramdisk they need > > to join the cluster > > and mount the shared root. Only potential master > > nodes require a local > > boot disk, because a potential master node can be > > the first node booted > > in the cluster. > >=20 > > In general, nodes 1 and 2 should be potential master > > nodes with local > > boot devices. The rest of the nodes do not need to > > be potential master > > nodes, and can use an Etherboot floppy. > >=20 > > Right now, all nodes need a swap device, and it must > > be the same device > > name on all nodes (e.g., /dev/hda3). I'd like to > > loosen that restriction > > in the future. Then you could have nodes with no > > internal disks, so long > > as they have enough memory to not need swap. > >=20 > > > In this case, when the machine gets > > > rebooted, we just save the floppy, or make lots of > > > copies for other machines to have handy ? > >=20 > > It's best to make a copy for each node, and keep > > that copy in the node's > > floppy drive. That way, the node can simply reboot > > and rejoin the > > cluster without manual intervention. > >=20 > > > Also, each node need a Fibre Channel HBA, but does > > the > > > lock server need an HBA ? > >=20 > > No. It's just a dumb lock server with no > > understanding of GFS. > >=20 > > > Thanks for helping out.. > >=20 > > No problem. > >=20 > > --=20 > > Brian Watson | "Now I don't know, but > > I been told it's > > Linux Kernel Developer | hard to run with the > > weight of gold, > > Open SSI Clustering Project | Other hand I heard it > > said, it's > > Compaq Computer Corp | just as hard with the > > weight of lead." > > Los Angeles, CA | -Robert Hunter, > > 1970 > >=20 > > mailto:Bri...@co... > > http://opensource.compaq.com/ >=20 >=20 > __________________________________________________ > Do You Yahoo!? > Yahoo! Movies - coverage of the 74th Academy Awards=AE > http://movies.yahoo.com/ >=20 > _______________________________________________ > ssic-linux-devel mailing list > ssi...@li... > https://lists.sourceforge.net/lists/listinfo/ssic-linux-devel |
From: John H. <john@Calva.COM> - 2002-03-21 08:28:09
|
> Right now, all nodes need a swap device, and it must be > the same device name on all nodes (e.g., /dev/hda3). I'd > like to loosen that restriction in the future. Then you > could have nodes with no internal disks, so long as they > have enough memory to not need swap. No swapping across the network? |
From: Bruce W. <br...@ka...> - 2002-03-21 19:04:19
|
> > No swapping across the network? > If/when there is a need, there are two different ways we could provide it, that I'm familiar with. There currently is remote device support in SSI, although the remote block device capability isn't completely there. Finishing that work and figuring out a way to configure it would give us swapping/paging to other nodes in the cluster, I believe. It may also be possible to do it HA, with device failover providing resiliency from a failure of the device node. The "problem" with the above approach is that each diskless client node needs it's own dedicated swap device (albeit all of them attached to one or two master-type nodes). NonStop Clusters for Unixware had a swap sharing capability, where one node could borrow swap space from another node if it ran out (even if it had some swap space locally). Providing this functionality on Linux is a low priority project, in my judgement. bruce > > _______________________________________________ > ssic-linux-devel mailing list > ssi...@li... > https://lists.sourceforge.net/lists/listinfo/ssic-linux-devel |