Re: [lustre-discuss] ZFS w/Lustre problem
On Wed, 18 Nov 2020, Faaland, Olaf P. wrote: You mentioned you may have a fix for zfs_send.c in ZFS. Although Lustre tickles the bug, it's not likely that is the only way to tickle it. Is there already a bug report for your issue at https://github.com/openzfs/zfs/issues? If not, can you create one, even if your patch isn't successful? That's the place to get your patch landed, and/or get help with the issue. The issue is already reported as #8067. The patch mentioned in this article is still valid, but for a different file; the txg_wait_synced() call is at line 2215 of dmu_send.c on ZFS 0.7.13. This patch does not fix that very slow performance of 'zfs recv', but it does fix the 'dataset does not exist' problem. Steve -- Steve Thompson E-mail: smt AT vgersoft DOT com Voyager Software LLC Web: http://www DOT vgersoft DOT com 3901 N Charles St VSW Support: support AT vgersoft DOT com Baltimore MD 21218 "186,282 miles per second: it's not just a good idea, it's the law" ___ lustre-discuss mailing list lustre-discuss@lists.lustre.org http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
Re: [lustre-discuss] ZFS w/Lustre problem
Hi Steve, You mentioned you may have a fix for zfs_send.c in ZFS. Although Lustre tickles the bug, it's not likely that is the only way to tickle it. Is there already a bug report for your issue at https://github.com/openzfs/zfs/issues? If not, can you create one, even if your patch isn't successful? That's the place to get your patch landed, and/or get help with the issue. thanks, -Olaf From: lustre-discuss on behalf of Steve Thompson Sent: Tuesday, November 10, 2020 5:06 AM To: Hans Henrik Happe Cc: lustre-discuss@lists.lustre.org Subject: Re: [lustre-discuss] ZFS w/Lustre problem On Mon, 9 Nov 2020, Hans Henrik Happe wrote: > I sounds like this issue, but I'm not sure what your dnodesize is: > > https://github.com/openzfs/zfs/issues/8458 > > ZFS 0.8.1+ on the receiving side should fix it. Then again ZFS 0.8 is > not supported in Lustre 2.12, so it's a bit hard to restore, without > copying the underlying devices. Hans Henrik, Many thanks for your input. I had in fact known about the dnodesize issue, and tested a workaround. Unfortunately, it turned out not to be this. Instead, I have tested a patch to zfs_send.c, which does appear to have solved the issue. The zfs send/recv is still running, however; if it completes successfully, I will post again with details of the patch. Steve -- Steve Thompson E-mail: smt AT vgersoft DOT com Voyager Software LLC Web: http://www DOT vgersoft DOT com 3901 N Charles St VSW Support: support AT vgersoft DOT com Baltimore MD 21218 "186,282 miles per second: it's not just a good idea, it's the law" ___ lustre-discuss mailing list lustre-discuss@lists.lustre.org http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org ___ lustre-discuss mailing list lustre-discuss@lists.lustre.org http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
Re: [lustre-discuss] ZFS w/Lustre problem
On Mon, 9 Nov 2020, Hans Henrik Happe wrote: I sounds like this issue, but I'm not sure what your dnodesize is: https://github.com/openzfs/zfs/issues/8458 ZFS 0.8.1+ on the receiving side should fix it. Then again ZFS 0.8 is not supported in Lustre 2.12, so it's a bit hard to restore, without copying the underlying devices. Hans Henrik, Many thanks for your input. I had in fact known about the dnodesize issue, and tested a workaround. Unfortunately, it turned out not to be this. Instead, I have tested a patch to zfs_send.c, which does appear to have solved the issue. The zfs send/recv is still running, however; if it completes successfully, I will post again with details of the patch. Steve -- Steve Thompson E-mail: smt AT vgersoft DOT com Voyager Software LLC Web: http://www DOT vgersoft DOT com 3901 N Charles St VSW Support: support AT vgersoft DOT com Baltimore MD 21218 "186,282 miles per second: it's not just a good idea, it's the law" ___ lustre-discuss mailing list lustre-discuss@lists.lustre.org http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
Re: [lustre-discuss] ZFS w/Lustre problem
I sounds like this issue, but I'm not sure what your dnodesize is: https://github.com/openzfs/zfs/issues/8458 ZFS 0.8.1+ on the receiving side should fix it. Then again ZFS 0.8 is not supported in Lustre 2.12, so it's a bit hard to restore, without copying the underlying devices. Cheers, Hans Henrik On 06.11.2020 21.23, Steve Thompson wrote: > This may be a question for the ZFS list... > > I have Lustre 2.12.5 on Centos 7.8 with ZFS 0.7.13, 10GB network. I > make snapshots of the Lustre filesystem with 'lctl snapshot_create' > and at a later time transfer these snapshots to a backup system with > zfs send/recv. This works well for everything but the MDT. For the > MDT, I find that the zfs recv always fails when a little less than 1GB > has been transferred (this being an incremental send/recv of snapshots > taken a day apart): > > # zfs send -v -c -i fs0pool/mdt0@03-nov-2020 fs0pool/mdt0@04-nov-2020 | \ > zfs recv -F backups/fs0pool/mdt0 > > 12:11:18 946M fs0pool/mdt0@04-nov-2020-01:00 > 12:11:19 946M fs0pool/mdt0@04-nov-2020-01:00 > 12:11:20 946M fs0pool/mdt0@04-nov-2020-01:00 > cannot receive incremental stream: dataset does not exist > > while if the data transfer is much smaller, the send/recv works. Since > once I get a failure it is not possible to complete a send/recv for > any subsequent day, I am doing a full snapshot send to a file; this > always works and takes about 5/6 minutes for my MDT. When using zfs > send/recv, the recv is always very very slow (several hours to get to > the above failure point, even when using mbuffer). I am using custom > zfs replication scripts, but it fails also using the zrep package. > > Does anyone know of a possible explanation? Is there any version of > ZFS 0.8 that works with Lustre 2.12.5? > > Thanks, > Steve ___ lustre-discuss mailing list lustre-discuss@lists.lustre.org http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
[lustre-discuss] ZFS w/Lustre problem
This may be a question for the ZFS list... I have Lustre 2.12.5 on Centos 7.8 with ZFS 0.7.13, 10GB network. I make snapshots of the Lustre filesystem with 'lctl snapshot_create' and at a later time transfer these snapshots to a backup system with zfs send/recv. This works well for everything but the MDT. For the MDT, I find that the zfs recv always fails when a little less than 1GB has been transferred (this being an incremental send/recv of snapshots taken a day apart): # zfs send -v -c -i fs0pool/mdt0@03-nov-2020 fs0pool/mdt0@04-nov-2020 | \ zfs recv -F backups/fs0pool/mdt0 12:11:18946M fs0pool/mdt0@04-nov-2020-01:00 12:11:19946M fs0pool/mdt0@04-nov-2020-01:00 12:11:20946M fs0pool/mdt0@04-nov-2020-01:00 cannot receive incremental stream: dataset does not exist while if the data transfer is much smaller, the send/recv works. Since once I get a failure it is not possible to complete a send/recv for any subsequent day, I am doing a full snapshot send to a file; this always works and takes about 5/6 minutes for my MDT. When using zfs send/recv, the recv is always very very slow (several hours to get to the above failure point, even when using mbuffer). I am using custom zfs replication scripts, but it fails also using the zrep package. Does anyone know of a possible explanation? Is there any version of ZFS 0.8 that works with Lustre 2.12.5? Thanks, Steve -- Steve Thompson E-mail: smt AT vgersoft DOT com Voyager Software LLC Web: http://www DOT vgersoft DOT com 3901 N Charles St VSW Support: support AT vgersoft DOT com Baltimore MD 21218 "186,282 miles per second: it's not just a good idea, it's the law" ___ lustre-discuss mailing list lustre-discuss@lists.lustre.org http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org