Re: [ceph-users] samba gateway experiences with cephfs ?
On Thu, 24 May 2018 15:13:09 +0200, Daniel Baumann wrote: > On 05/24/2018 02:53 PM, David Disseldorp wrote: > >> [ceph_test] > >> path = /ceph-kernel > >> guest ok = no > >> delete readonly = yes > >> oplocks = yes > >> posix locking = no > > jftr, we use the following to disable all locking (on samba 4.8.2): > > oplocks = False > level2 oplocks = False > kernel oplocks = no oplocks aren't locks per se - they allow the client to cache data locally (leases in SMB2+), often allowing for improved application performance. That said, if the same share path is accessible via NFS or native CephFS then oplocks / leases should be disabled, until proper vfs_ceph lease support is implemented via the delegation API. Cheers, David ___ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
Re: [ceph-users] samba gateway experiences with cephfs ?
Hi David, Many thanks for your help :) I'm using Scientific Linux 7.5 thus samba-4.7.1-6.el7.x86_64 I've added these settings to the share: aio read size = 1 aio write size = 1 ...and restarting samba, Helios LanTest didn't show any real changes, I will test from a Linux machine later and see if I/O improves here. Glad to hear CTDB will work with posix locks off, I will start testing this next week. Oh, the RADOS object lock is definitely worth investigating... thanks for this too :) all the best, Jake On 24/05/18 13:53, David Disseldorp wrote: > Hi Jake, > > On Thu, 24 May 2018 13:17:16 +0100, Jake Grimmett wrote: > >> Hi Daniel, David, >> >> Many thanks for both of your advice. >> >> Sorry not to reply to the list, but I'm subscribed to the digest and my >> mail client will not reply to individual threads - I've switched back to >> regular. > > No worries, cc'ing the list in this response. > >> As to this issue, I've turned off posix locking, which has improved >> write speeds - here are the old benchmarks plus new figures. >> >> i.e. Using Helios LanTest 6.0.0 on Osx. >> >> Create 300 Files >> Cephfs (kernel) > samba (no Posix locks) >> average 3600 ms >> Cephfs (kernel) > samba. average 5100 ms >> Isilon > CIFS average 2600 ms >> ZFS > samba average 121 ms >> >> Remove 300 files >> Cephfs (kernel) > samba (no Posix locks) >> average 2200 ms >> Cephfs (kernel) > samba. average 2100 ms >> Isilon > CIFS average 900 ms >> ZFS > samba average 421 ms >> >> Write 300MB to file >> Cephfs (kernel) > samba (no Posix locks) >> average 53 MB/s >> Cephfs (kernel) > samba. average 25 MB/s >> Isilon > CIFS average 17.9 MB/s >> ZFS > samba average 64.4 MB/s >> >> >> Settings as follows: >> [global] >> (snip) >> smb2 leases = yes >> >> >> [ceph_test] >> path = /ceph-kernel >> guest ok = no >> delete readonly = yes >> oplocks = yes >> posix locking = no > > Which version of Samba are you using here? If it's relatively recent > (4.6+), please rerun with asynchronous I/O enabled via: > [share] > aio read size = 1 > aio write size = 1 > > ...these settings are the default with Samba 4.8+. AIO won't help the > file creation / deletion benchmarks, but there should be a positive > affect on read/write performance. > >> Disabling all locking (locking = no) gives some further speed improvements. >> >> File locking hopefully will not be an issue... >> >> We are not exporting this share via NFS. The shares will only be used by >> single clients (Windows or OSX Desktops) as a backup location. >> >> Specifically, each machine has a separate smb mounted folder, to which >> they either use ChronoSync or Max SyncUp to write to. >> >> One other point... >> Will CTDB work with "posix locking = no"? >> It would be great if CTDB works, as I'd like to have a several SMB heads >> to load-balance the clients > > Yes, it shouldn't affect CTDB. Clustered FS POSIX locks are used by CTDB > for split-brain avoidance, and are separate to Samba's > client-lock <-> POSIX-lock mapping. > (https://wiki.samba.org/index.php/Configuring_the_CTDB_recovery_lock) > FYI, CTDB is now also capable of using RADOS objects for the recovery > lock: > https://ctdb.samba.org/manpages/ctdb_mutex_ceph_rados_helper.7.html > > Cheers, David > ___ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
Re: [ceph-users] samba gateway experiences with cephfs ?
Hi, On 05/24/2018 02:53 PM, David Disseldorp wrote: >> [ceph_test] >> path = /ceph-kernel >> guest ok = no >> delete readonly = yes >> oplocks = yes >> posix locking = no jftr, we use the following to disable all locking (on samba 4.8.2): oplocks = False level2 oplocks = False kernel oplocks = no Regards, Daniel ___ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
Re: [ceph-users] samba gateway experiences with cephfs ?
Hi Jake, On Thu, 24 May 2018 13:17:16 +0100, Jake Grimmett wrote: > Hi Daniel, David, > > Many thanks for both of your advice. > > Sorry not to reply to the list, but I'm subscribed to the digest and my > mail client will not reply to individual threads - I've switched back to > regular. No worries, cc'ing the list in this response. > As to this issue, I've turned off posix locking, which has improved > write speeds - here are the old benchmarks plus new figures. > > i.e. Using Helios LanTest 6.0.0 on Osx. > > Create 300 Files > Cephfs (kernel) > samba (no Posix locks) >average 3600 ms > Cephfs (kernel) > samba. average 5100 ms > Isilon > CIFS average 2600 ms > ZFS > samba average 121 ms > > Remove 300 files > Cephfs (kernel) > samba (no Posix locks) >average 2200 ms > Cephfs (kernel) > samba. average 2100 ms > Isilon > CIFS average 900 ms > ZFS > samba average 421 ms > > Write 300MB to file > Cephfs (kernel) > samba (no Posix locks) >average 53 MB/s > Cephfs (kernel) > samba. average 25 MB/s > Isilon > CIFS average 17.9 MB/s > ZFS > samba average 64.4 MB/s > > > Settings as follows: > [global] > (snip) > smb2 leases = yes > > > [ceph_test] > path = /ceph-kernel > guest ok = no > delete readonly = yes > oplocks = yes > posix locking = no Which version of Samba are you using here? If it's relatively recent (4.6+), please rerun with asynchronous I/O enabled via: [share] aio read size = 1 aio write size = 1 ...these settings are the default with Samba 4.8+. AIO won't help the file creation / deletion benchmarks, but there should be a positive affect on read/write performance. > Disabling all locking (locking = no) gives some further speed improvements. > > File locking hopefully will not be an issue... > > We are not exporting this share via NFS. The shares will only be used by > single clients (Windows or OSX Desktops) as a backup location. > > Specifically, each machine has a separate smb mounted folder, to which > they either use ChronoSync or Max SyncUp to write to. > > One other point... > Will CTDB work with "posix locking = no"? > It would be great if CTDB works, as I'd like to have a several SMB heads > to load-balance the clients Yes, it shouldn't affect CTDB. Clustered FS POSIX locks are used by CTDB for split-brain avoidance, and are separate to Samba's client-lock <-> POSIX-lock mapping. (https://wiki.samba.org/index.php/Configuring_the_CTDB_recovery_lock) FYI, CTDB is now also capable of using RADOS objects for the recovery lock: https://ctdb.samba.org/manpages/ctdb_mutex_ceph_rados_helper.7.html Cheers, David ___ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
Re: [ceph-users] samba gateway experiences with cephfs ?
Hi Daniel and Jake, On Mon, 21 May 2018 22:46:01 +0200, Daniel Baumann wrote: > Hi > > On 05/21/2018 05:38 PM, Jake Grimmett wrote: > > Unfortunately we have a large number (~200) of Windows and Macs clients > > which need CIFS/SMB access to cephfs. > > we too, which is why we're (partially) exporting cephfs over samba too, > 1.5y in production now. > > for us, cephfs-over-samba is significantly slower than cephfs directly > too, but it's not really an issue here (basically, if people use a > windows client here, they're already on the slow track anyway). > > we had to do two things to get it working reliably though: > > a) disable all locking on samba (otherwise "opportunistic locking" on > windows clients killed within hours all mds (kraken at that time)) Have you seen this on more recent versions? Please raise a bug if so - client induced MDS outages would be a pretty serious issue. If your share path is isolated from non-samba clients (as you mention below), then allowing clients to cache reads / writes locally via oplocks / SMB2+ leases should offer a significant performance improvements. > b) only allow writes to a specific space on cephfs, reserved to samba > (with luminous; otherwise, we'd have problems with data consistency on > cephfs with people writing the same files from linux->cephfs and > samba->cephfs concurrently). my hunch is that samba caches writes and > doesn't give them back appropriatly. If you're sharing a kernel CephFS mount, then the Linux page cache will be used for Samba share I/O, but Samba will fully abide by client sync requests if "strict sync = yes" (default in Samba 4.7+). > > Finally, is the vfs_ceph module for Samba useful? It doesn't seem to be > > widely available pre-complied for for RHEL derivatives. Can anyone > > comment on their experiences using vfs_ceph, or point me to a Centos 7.x > > repo that has it? > > we use debian, with backported kernel and backported samba, which has > vfs_ceph pre-compiled. however, we couldn't make vfs_ceph work at all - > the snapshot patters just don't seem to match/align (and nothing we > tried seem to work). vfs_ceph doesn't support snapshots at this stage. I hope to work on this feature in the near future, so that it's fully integrated with the Explorer previous versions UI. Cheers, David ___ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
Re: [ceph-users] samba gateway experiences with cephfs ?
Hi On 05/21/2018 05:38 PM, Jake Grimmett wrote: > Unfortunately we have a large number (~200) of Windows and Macs clients > which need CIFS/SMB access to cephfs. we too, which is why we're (partially) exporting cephfs over samba too, 1.5y in production now. for us, cephfs-over-samba is significantly slower than cephfs directly too, but it's not really an issue here (basically, if people use a windows client here, they're already on the slow track anyway). we had to do two things to get it working reliably though: a) disable all locking on samba (otherwise "opportunistic locking" on windows clients killed within hours all mds (kraken at that time)) b) only allow writes to a specific space on cephfs, reserved to samba (with luminous; otherwise, we'd have problems with data consistency on cephfs with people writing the same files from linux->cephfs and samba->cephfs concurrently). my hunch is that samba caches writes and doesn't give them back appropriatly. > Finally, is the vfs_ceph module for Samba useful? It doesn't seem to be > widely available pre-complied for for RHEL derivatives. Can anyone > comment on their experiences using vfs_ceph, or point me to a Centos 7.x > repo that has it? we use debian, with backported kernel and backported samba, which has vfs_ceph pre-compiled. however, we couldn't make vfs_ceph work at all - the snapshot patters just don't seem to match/align (and nothing we tried seem to work). Regards, Daniel ___ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
[ceph-users] samba gateway experiences with cephfs ?
Dear All, Excited to see snapshots finally becoming a stable feature in cephfs :) Unfortunately we have a large number (~200) of Windows and Macs clients which need CIFS/SMB access to cephfs. None-the-less, snapshots have prompted us to start testing ceph to see if we can use it as a scale-out NAS... cephfs native performance on our test setup appears good, however tests accessing via samba have been slightly disappointing, especially with small file I/O. Large file I/O is fair, but could still be improved. Using Helios LanTest 6.0.0 on Osx. Create 300 Files Cephfs (kernel) > samba. average 5100 ms Isilon > CIFS average 2600 ms ZFS > samba average 121 ms Remove 300 files Cephfs (kernel) > samba. average 2100 ms Isilon > CIFS average 900 ms ZFS > samba average 421 ms Write 300MB to file Cephfs (kernel) > samba. average 25 MB/s Isilon > CIFS average 17.9 MB/s ZFS > samba average 64.4 MB/s Hardware Used: CephFS: five node dual Xeon cluster (120 bluestore OSD, 4 x nvme metadata for Cephfs, bulk data EC 4+1), Scientific Linux 7.5, ceph 12.2.5, kernel client (fuse significantly slower). Isilon: 6 year old, 8 x NL108 ZFS: SL 6.4 on a Dell R730XD, 24 x 1.8TB drives Ceph Samba gateway is a separate machine: dual Xeon, 40Gb ethernet, 128GB RAM, also running SL 7.5. Finally, is the vfs_ceph module for Samba useful? It doesn't seem to be widely available pre-complied for for RHEL derivatives. Can anyone comment on their experiences using vfs_ceph, or point me to a Centos 7.x repo that has it? many thanks for all and any advice, Jake ___ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com