On Mon, Sep 17, 2018 at 5:39 AM, Jeffrey Zhang <zhang.lei.fly@xxxxxxxxx> wrote: > In one env, which is deployed through container, i found the ceph-osd always > be suicide due to "error (24) Too many open files" > > Then i increased the LimitNOFILE for the container from 65k to 655k, which > could fix the issue. > But the FDs increase all the time. now the max number is around 155k. I am > afraid > it will increase forever. > > I also found there is a option `max_open_files`, but it seems only used for > upstart scripts at OS level, and the default value is > 16384 now[2]. whereas if you are using systemd, it will never load from the > `max_open_files` options and fixed to 1048576 in default[1]. > So i guess if the ceph-osd live long enough, it will still read the OS level > strict and suicide at the end. > > So here is the question > 1. since almost all os distro already moved to systemd, so max_open_files is > uselss now. > 2. is there any mechanism that ceph-osd could release some fds? If an OSD actually hits its file limits it just sort of stops; we don't have mechanisms to trim them down through the whole process. This is because we need a socket descriptor for every connection (which you can consider to be every client, plus a monitor and manager, plus a multiple on the number of OSD peers it has), plus the files open (which mostly only applies to FileStore and there's a config, defaults to 1024 I think). 155k is large but I've seen that number before. In systemd we just set it to 1 million at this point, which has not come up as a limit for anybody yet. :) -Greg > > > [0] > https://github.com/ceph/ceph/commit/16c603b26f3d7dfcf0028e17fe3c04d94a434387 > [1] https://github.com/ceph/ceph/commit/8453a89 > [2] > https://github.com/ceph/ceph/commit/672c56b18de3b02606e47013edfc2e8b679d8797 > > > > -- > Regards, > Jeffrey Zhang > Blog: http://xcodest.me > > _______________________________________________ > ceph-users mailing list > ceph-users@xxxxxxxxxxxxxx > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >
