Re: [ovirt-devel] [ovirt-users] Recommended setup for a FC based storage domain
CC'ing the devel list, maybe some VDSM and storage people can explain this? Am 10.06.2014 12:24, schrieb combuster:
/etc/libvirt/libvirtd.conf and /etc/vdsm/logger.conf
, but unfortunately maybe I've jumped to conclusions, last weekend, that very same thin provisioned vm was running a simple export for 3hrs before I've killed the process. But I wondered:
1. The process that runs behind the export is qemu-img convert (from raw to raw), and running iotop shows that every three or four seconds it reads 10-13 MBps and then idles for a few seconds. Run the numbers on 100GB (why is he covering the entire 100 of 15GB used on thin volume I still don't get it) and you get precisely 3-4 hrs estimated time remaining. 2. When I run export with SPM on a node that doesn't have any vm's running, export finishes for aprox. 30min (iotop shows 40-70MBps read speed constantly) 3. Renicing I/O priority of the qemu-img process as well as the CPU priority gave no results, it was still runing slow beyond any explanation.
Debug logs showed nothing of interest, so I disabled anything above warning and it suddenly accelerated the export, so I've connected the wrong dots.
On 06/10/2014 11:18 AM, Andrew Lau wrote:
Interesting, which files did you modify to lower the log levels?
On Tue, Jun 3, 2014 at 12:38 AM, <combuster@archlinux.us> wrote:
One word of caution so far, when exporting any vm, the node that acts as SPM is stressed out to the max. I releived the stress by a certain margin with lowering libvirtd and vdsm log levels to WARNING. That shortened out the export procedure by at least five times. But vdsm process on the SPM node is still with high cpu usage so it's best that the SPM node should be left with a decent CPU time amount to spare. Also, export of VM's with high vdisk capacity and thin provisioning enabled (let's say 14GB used of 100GB defined) took around 50min over a 10Gb ethernet interface to a 1Gb export NAS device that was not stressed out at all by other processes. When I did that export with debug log levels it took 5hrs :(
So lowering log levels is a must in production enviroment. I've deleted the lun that I exported on the storage (removed it first from ovirt) and for the next weekend I am planing to add a new one, export it again on all the nodes and start a few fresh vm installations. Things I'm going to look for are partition alignment and running them from different nodes in the cluster at the same time. I just hope that not all I/O is going to pass through the SPM, this is the one thing that bothers me the most.
I'll report back on these results next week, but if anyone has experience with this kind of things or can point to some documentation would be great.
On Monday, 2. June 2014. 18.51.52 you wrote:
I'm curious to hear what other comments arise, as we're analyzing a production setup shortly.
On Sun, Jun 1, 2014 at 10:11 PM, <combuster@archlinux.us> wrote:
I need to scratch gluster off because setup is based on CentOS 6.5, so essential prerequisites like qemu 1.3 and libvirt 1.0.1 are not met. Gluster would still work with EL6, afaik it just won't use libgfapi and instead use just a standard mount.
Any info regarding FC storage domain would be appreciated though.
Thanks
Ivan
On Sunday, 1. June 2014. 11.44.33 combuster@archlinux.us wrote:
Hi,
I have a 4 node cluster setup and my storage options right now are a FC based storage, one partition per node on a local drive (~200GB each) and a NFS based NAS device. I want to setup export and ISO domain on the NAS and there are no issues or questions regarding those two. I wasn't aware of any other options at the time for utilizing a local storage (since this is a shared based datacenter) so I exported a directory from each partition via NFS and it works. But I am little in the dark with the following:
1. Are there any advantages for switching from NFS based local storage to a Gluster based domain with blocks for each partition. I guess it can be only performance wise but maybe I'm wrong. If there are advantages, are there any tips regarding xfs mount options etc ?
2. I've created a volume on the FC based storage and exported it to all of the nodes in the cluster on the storage itself. I've configured multipathing correctly and added an alias for the wwid of the LUN so I can distinct this one and any other future volumes more easily. At first I created a partition on it but since oVirt saw only the whole LUN as raw device I erased it before adding it as the FC master storage domain. I've imported a few VM's and point them to the FC storage domain. This setup works, but:
- All of the nodes see a device with the alias for the wwid of the volume, but only the node wich is currently the SPM for the cluster can see logical volumes inside. Also when I setup the high availability for VM's residing on the FC storage and select to start on any node on the cluster, they always start on the SPM. Can multiple nodes run different VM's on the same FC storage at the same time (logical thing would be that they can, but I wanted to be sure first). I am not familiar with the logic oVirt utilizes that locks the vm's logical volume to prevent corruption.
- Fdisk shows that logical volumes on the LUN of the FC volume are missaligned (partition doesn't end on cylindar boundary), so I wonder if this is becuase I imported the VM's with disks that were created on local storage before and that any _new_ VM's with disks on the fc storage would be propperly aligned.
This is a new setup with oVirt 3.4 (did an export of all the VM's on 3.3 and after a fresh installation of the 3.4 imported them back again). I have room to experiment a little with 2 of the 4 nodes because currently they are free from running any VM's, but I have limited room for anything else that would cause an unplanned downtime for four virtual machines running on the other two nodes on the cluster (currently highly available and their drives are on the FC storage domain). All in all I have 12 VM's running and I'm asking on the list for advice and guidance before I make any changes.
Just trying to find as much info regarding all of this as possible before acting upon.
Thank you in advance,
Ivan
-- Mit freundlichen Grüßen / Regards Sven Kieske Systemadministrator Mittwald CM Service GmbH & Co. KG Königsberger Straße 6 32339 Espelkamp T: +49-5772-293-100 F: +49-5772-293-333 https://www.mittwald.de Geschäftsführer: Robert Meyer St.Nr.: 331/5721/1033, USt-IdNr.: DE814773217, HRA 6640, AG Bad Oeynhausen Komplementärin: Robert Meyer Verwaltungs GmbH, HRB 13260, AG Bad Oeynhausen
Thanks, I really hope someone can help, because right now I'm afraid to thin provision any large volume due to this. I forgot to mention that, when I do the export on the NAS, node that is SPM at that moment is running qemu-img convert process and VDSM process on that node is running wild (400% cpu load from time to time), but qemu-img convert never occupies more than 1 thread (100%). There are a few (4-5) nfsd processes that are going from 40% to 90%. CPU load on the engine admin panel shows that vm's that run on SPM show abnormaly high cpu load (40-60%) and they usually run idle at the time. So: - Runing exprort when SPM node does not have any vm's running is fast (thin or fully provisioned, doesn't matteR) - Running export when SPM node does have 3-4 vm's running is painfully slow for thin provisioned vm's - Running export when SPM node does have 3-4 vm's running but vm that is exporting is not thin provisioned is fast enough even though vm's on SPM show an increased cpu load also I'm yet to measure I/O utilization in all scenarios, but I'm positive on what I wrote for the thin provisioned volume when there are vms runing on spm, it goes 15MBps max and just at bursts every three or four seconds (I measure this on the SPM node because the qemu-img convert runs on that node, even though vm is residing on another node, with its thin provisioned disk on its local storage shared via nfs). On čet, 2014-06-12 at 13:31 +0000, Sven Kieske wrote:
CC'ing the devel list, maybe some VDSM and storage people can explain this?
Am 10.06.2014 12:24, schrieb combuster:
/etc/libvirt/libvirtd.conf and /etc/vdsm/logger.conf
, but unfortunately maybe I've jumped to conclusions, last weekend, that very same thin provisioned vm was running a simple export for 3hrs before I've killed the process. But I wondered:
1. The process that runs behind the export is qemu-img convert (from raw to raw), and running iotop shows that every three or four seconds it reads 10-13 MBps and then idles for a few seconds. Run the numbers on 100GB (why is he covering the entire 100 of 15GB used on thin volume I still don't get it) and you get precisely 3-4 hrs estimated time remaining. 2. When I run export with SPM on a node that doesn't have any vm's running, export finishes for aprox. 30min (iotop shows 40-70MBps read speed constantly) 3. Renicing I/O priority of the qemu-img process as well as the CPU priority gave no results, it was still runing slow beyond any explanation.
Debug logs showed nothing of interest, so I disabled anything above warning and it suddenly accelerated the export, so I've connected the wrong dots.
On 06/10/2014 11:18 AM, Andrew Lau wrote:
Interesting, which files did you modify to lower the log levels?
On Tue, Jun 3, 2014 at 12:38 AM, <combuster@archlinux.us> wrote:
One word of caution so far, when exporting any vm, the node that acts as SPM is stressed out to the max. I releived the stress by a certain margin with lowering libvirtd and vdsm log levels to WARNING. That shortened out the export procedure by at least five times. But vdsm process on the SPM node is still with high cpu usage so it's best that the SPM node should be left with a decent CPU time amount to spare. Also, export of VM's with high vdisk capacity and thin provisioning enabled (let's say 14GB used of 100GB defined) took around 50min over a 10Gb ethernet interface to a 1Gb export NAS device that was not stressed out at all by other processes. When I did that export with debug log levels it took 5hrs :(
So lowering log levels is a must in production enviroment. I've deleted the lun that I exported on the storage (removed it first from ovirt) and for the next weekend I am planing to add a new one, export it again on all the nodes and start a few fresh vm installations. Things I'm going to look for are partition alignment and running them from different nodes in the cluster at the same time. I just hope that not all I/O is going to pass through the SPM, this is the one thing that bothers me the most.
I'll report back on these results next week, but if anyone has experience with this kind of things or can point to some documentation would be great.
On Monday, 2. June 2014. 18.51.52 you wrote:
I'm curious to hear what other comments arise, as we're analyzing a production setup shortly.
On Sun, Jun 1, 2014 at 10:11 PM, <combuster@archlinux.us> wrote:
I need to scratch gluster off because setup is based on CentOS 6.5, so essential prerequisites like qemu 1.3 and libvirt 1.0.1 are not met. Gluster would still work with EL6, afaik it just won't use libgfapi and instead use just a standard mount.
Any info regarding FC storage domain would be appreciated though.
Thanks
Ivan
On Sunday, 1. June 2014. 11.44.33 combuster@archlinux.us wrote: > Hi, > > I have a 4 node cluster setup and my storage options right now are > a FC > based storage, one partition per node on a local drive (~200GB > each) and > a > NFS based NAS device. I want to setup export and ISO domain on the > NAS > and > there are no issues or questions regarding those two. I wasn't > aware of > any > other options at the time for utilizing a local storage (since > this is a > shared based datacenter) so I exported a directory from each > partition > via > NFS and it works. But I am little in the dark with the following: > > 1. Are there any advantages for switching from NFS based local > storage to > a > Gluster based domain with blocks for each partition. I guess it > can be > only > performance wise but maybe I'm wrong. If there are advantages, are > there > any tips regarding xfs mount options etc ? > > 2. I've created a volume on the FC based storage and exported it > to all > of > the nodes in the cluster on the storage itself. I've configured > multipathing correctly and added an alias for the wwid of the LUN > so I > can > distinct this one and any other future volumes more easily. At > first I > created a partition on it but since oVirt saw only the whole LUN > as raw > device I erased it before adding it as the FC master storage > domain. I've > imported a few VM's and point them to the FC storage domain. This > setup > works, but: > > - All of the nodes see a device with the alias for the wwid of the > volume, > but only the node wich is currently the SPM for the cluster can see > logical > volumes inside. Also when I setup the high availability for VM's > residing > on the FC storage and select to start on any node on the cluster, > they > always start on the SPM. Can multiple nodes run different VM's on the > same > FC storage at the same time (logical thing would be that they can, > but I > wanted to be sure first). I am not familiar with the logic oVirt > utilizes > that locks the vm's logical volume to prevent corruption. > > - Fdisk shows that logical volumes on the LUN of the FC volume are > missaligned (partition doesn't end on cylindar boundary), so I > wonder if > this is becuase I imported the VM's with disks that were created > on local > storage before and that any _new_ VM's with disks on the fc > storage would > be propperly aligned. > > This is a new setup with oVirt 3.4 (did an export of all the VM's > on 3.3 > and after a fresh installation of the 3.4 imported them back > again). I > have room to experiment a little with 2 of the 4 nodes because > currently > they are free from running any VM's, but I have limited room for > anything else that would cause an unplanned downtime for four virtual > machines running on the other two nodes on the cluster (currently > highly > available and their drives are on the FC storage domain). All in > all I > have 12 VM's running and I'm asking on the list for advice and > guidance > before I make any changes. > > Just trying to find as much info regarding all of this as possible > before > acting upon. > > Thank you in advance, > > Ivan
-- Mit freundlichen Grüßen / Regards
Sven Kieske
Systemadministrator Mittwald CM Service GmbH & Co. KG Königsberger Straße 6 32339 Espelkamp T: +49-5772-293-100 F: +49-5772-293-333 https://www.mittwald.de Geschäftsführer: Robert Meyer St.Nr.: 331/5721/1033, USt-IdNr.: DE814773217, HRA 6640, AG Bad Oeynhausen Komplementärin: Robert Meyer Verwaltungs GmbH, HRB 13260, AG Bad Oeynhausen
participants (2)
-
Ivan Bulatovic -
Sven Kieske