Hi Alexander, it might be that you are expecting too much from ceph. The design of the filesystem was not some grand plan with every detail worked out. It was more the classic evolutionary approach, something working was screwed on top of rados and things evolved from there on. It is possible that the code and pin-seperation is not as clean as one would imagine. Here is what I observe before and after pinning everything explicitly: - before pinning: * high MDS load for no apparent reason - the balancer was just going in circles * stopping an MDS would besically bring all IO down - after pinning: * low MDS load, better user performance, much faster restarts * stopping an MDS does not kill all IO immediately, some IO continues, however, eventually every client gets stuck There is apparently still communication between all ranks about all clients and it is a bit annoying that some of this communication is blocking. Not sure if it has to be blocking or if one could make it asynchronous requests to the down rank. My impression is that ceph internals are rather bad at making stuff asynchronous. So if something in the MDS cluster is not healthy sooner or later IO will stop waiting for some blocking request to the unhealthy MDS. There seems to be no such thing as IO on other healthy MDSes continues as usual. Specifically rank 0 is critical. Best regards, ================= Frank Schilder AIT Risø Campus Bygning 109, rum S14 ________________________________________ From: Eugen Block <eblock@xxxxxx> Sent: Thursday, November 21, 2024 9:36 AM To: Александр Руденко Cc: ceph-users@xxxxxxx Subject: Re: [CephFS] Completely exclude some MDS rank from directory processing I'm not aware of any hard limit for the number of Filesystems, but that doesn't really mean very much. IIRC, last week during a Clyso talk at Eventbrite I heard someone say that they deployed around 200 Filesystems or so, I don't remember if it was a production environment or just a lab environment. I assume that you would probably be limited by the number of OSDs/PGs rather than by the number of Filesystems, 200 Filesystems require at least 400 pools. But maybe someone else has more experience in scaling CephFS that way. What we did was to scale the number of active MDS daemons for one CephFS. I believe in the end the customer had 48 MDS daemons on three MDS servers, 16 of them were active with directory pinning, at that time they had 16 standby-replay and 16 standby daemons. But it turned out that standby-replay didn't help their use case, so we disabled standby-replay. Can you show the entire 'ceph fs status' output? Any maybe also 'ceph fs dump'? Zitat von Александр Руденко <a.rudikk@xxxxxxxxx>: >> >> Just for testing purposes, have you tried pinning rank 1 to some other >> directory? Does it still break the CephFS if you stop it? > > > Yes, nothing changed. > > It's no problem that FS hangs when one of the ranks goes down, we will have > standby-reply for all ranks. I don't like that rank which is not pinned to > some dir handled some io of this dir or from clients which work with this > dir. > I mean that I can't robustly and fully separate client IO by ranks. > > Would it be an option to rather use multiple Filesystems instead of >> multi-active for one CephFS? > > > Yes, it's an option. But it is much more complicated in our case. Btw, do > you know how many different FS can be created in one cluster? Maybe you > know some potential problems with 100-200 FSs in one cluster? > > ср, 20 нояб. 2024 г. в 17:50, Eugen Block <eblock@xxxxxx>: > >> Ah, I misunderstood, I thought you wanted an even distribution across >> both ranks. >> Just for testing purposes, have you tried pinning rank 1 to some other >> directory? Does it still break the CephFS if you stop it? I'm not sure >> if you can prevent rank 1 from participating, I haven't looked into >> all the configs in quite a while. Would it be an option to rather use >> multiple Filesystems instead of multi-active for one CephFS? >> >> Zitat von Александр Руденко <a.rudikk@xxxxxxxxx>: >> >> > No it's not a typo. It's misleading example) >> > >> > dir1 and dir2 are pinned to rank 0, but FS and dir1,dir2 can't work >> without >> > rank 1. >> > rank 1 is used for something when I work with this dirs. >> > >> > ceph 16.2.13, metadata balancer and policy based balancing not used. >> > >> > ср, 20 нояб. 2024 г. в 16:33, Eugen Block <eblock@xxxxxx>: >> > >> >> Hi, >> >> >> >> > After pinning: >> >> > setfattr -n ceph.dir.pin -v 0 /fs-mountpoint/dir1 >> >> > setfattr -n ceph.dir.pin -v 0 /fs-mountpoint/dir2 >> >> >> >> is this a typo? If not, you did pin both directories to the same rank. >> >> >> >> Zitat von Александр Руденко <a.rudikk@xxxxxxxxx>: >> >> >> >> > Hi, >> >> > >> >> > I try to distribute all top level dirs in CephFS by different MDS >> ranks. >> >> > I have two active MDS with rank *0* and *1 *and I have 2 top dirs like >> >> > */dir1* and* /dir2*. >> >> > >> >> > After pinning: >> >> > setfattr -n ceph.dir.pin -v 0 /fs-mountpoint/dir1 >> >> > setfattr -n ceph.dir.pin -v 0 /fs-mountpoint/dir2 >> >> > >> >> > I can see next INOS and DNS distribution: >> >> > RANK STATE MDS ACTIVITY DNS INOS DIRS CAPS >> >> > 0 active c Reqs: 127 /s 12.6k 12.5k 333 505 >> >> > 1 active b Reqs: 11 /s 21 24 19 1 >> >> > >> >> > When I write to dir1 I can see a small amount on Reqs: in rank 1. >> >> > >> >> > Events in journal of MDS with rank 1: >> >> > cephfs-journal-tool --rank=fs1:1 event get list >> >> > >> >> > 2024-11-20T12:24:42.045056+0300 0xc5c1cb UPDATE: >> (scatter_writebehind) >> >> > A2037D53 >> >> > 2024-11-20T12:24:46.935934+0300 0xc5c629 SESSION: () >> >> > 2024-11-20T12:24:47.192012+0300 0xc5c7cd UPDATE: (lock inest >> accounted >> >> > scatter stat update) >> >> > 2024-11-20T12:24:47.904717+0300 0xc5ca0b SESSION: () >> >> > 2024-11-20T12:26:46.912719+0300 0xc5ca98 SESSION: () >> >> > 2024-11-20T12:26:47.910806+0300 0xc5cc3c SESSION: () >> >> > 2024-11-20T12:27:35.746239+0300 0xc5ccc9 SESSION: () >> >> > 2024-11-20T12:28:46.923812+0300 0xc5ce63 SESSION: () >> >> > 2024-11-20T12:28:47.903066+0300 0xc5d007 SESSION: () >> >> > 2024-11-20T12:29:08.063326+0300 0xc5d094 EXPORT: () >> >> > di1/A2037D53 >> >> > 2024-11-20T12:30:46.909621+0300 0xc5d96f SESSION: () >> >> > 2024-11-20T12:30:47.908050+0300 0xc5db13 SESSION: () >> >> > 2024-11-20T12:32:46.907649+0300 0xc5dba0 SESSION: () >> >> > 2024-11-20T12:32:47.905962+0300 0xc5dd44 SESSION: () >> >> > 2024-11-20T12:34:44.349348+0300 0xc5ddd1 SESSIONS: () >> >> > >> >> > But the main problem, when I stop MDS rank 1 (without any kind of >> >> standby) >> >> > - FS hangs for all actions. >> >> > Is this correct? Is it possible to completely exclude rank 1 from >> >> > processing dir1 and not stop io when rank 1 goes down? >> >> > _______________________________________________ >> >> > ceph-users mailing list -- ceph-users@xxxxxxx >> >> > To unsubscribe send an email to ceph-users-leave@xxxxxxx >> >> >> >> >> >> _______________________________________________ >> >> ceph-users mailing list -- ceph-users@xxxxxxx >> >> To unsubscribe send an email to ceph-users-leave@xxxxxxx >> >> >> >> >> >> _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
 |