Re: MDS performance issue

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, Jan 23, 2019 at 10:02 AM Albert Yue <transuranium.yue@xxxxxxxxx> wrote:
>
> But with enough memory on MDS, I can just cache all metadata into memory. Right now there are around 500GB metadata in the ssd. So this is not enough?
>

mds needs to tracking lots of extra information for each object. For
500G metadata, mds may need 1T or more memory.

> On Tue, Jan 22, 2019 at 5:48 PM Yan, Zheng <ukernel@xxxxxxxxx> wrote:
>>
>> On Tue, Jan 22, 2019 at 10:49 AM Albert Yue <transuranium.yue@xxxxxxxxx> wrote:
>> >
>> > Hi Yan Zheng,
>> >
>> > In your opinion, can we resolve this issue by move MDS to a 512GB or 1TB memory machine?
>> >
>>
>> The problem is from client side, especially clients with large memory.
>> I don't think enlarge mds cache size is good idea. you can
>> periodically check periodically
>> each kernel clients' /sys/kernel/debug/ceph/xxx/caps. run 'echo 2
>> >/proc/sys/vm/drop_caches' if a client used too many caps (for example
>> 10k),
>>
>> > On Mon, Jan 21, 2019 at 10:49 PM Yan, Zheng <ukernel@xxxxxxxxx> wrote:
>> >>
>> >> On Mon, Jan 21, 2019 at 11:16 AM Albert Yue <transuranium.yue@xxxxxxxxx> wrote:
>> >> >
>> >> > Dear Ceph Users,
>> >> >
>> >> > We have set up a cephFS cluster with 6 osd machines, each with 16 8TB harddisk. Ceph version is luminous 12.2.5. We created one data pool with these hard disks and created another meta data pool with 3 ssd. We created a MDS with 65GB cache size.
>> >> >
>> >> > But our users are keep complaining that cephFS is too slow. What we observed is cephFS is fast when we switch to a new MDS instance, once the cache fills up (which will happen very fast), client became very slow when performing some basic filesystem operation such as `ls`.
>> >> >
>> >>
>> >> It seems that clients hold lots of unused inodes their icache, which
>> >> prevent mds from trimming corresponding objects from its cache.  mimic
>> >> has command "ceph daemon mds.x cache drop" to ask client to drop its
>> >> cache. I'm also working on a patch that make kclient client release
>> >> unused inodes.
>> >>
>> >> For luminous,  there is not much we can do, except periodically run
>> >> "echo 2 > /proc/sys/vm/drop_caches"  on each client.
>> >>
>> >>
>> >> > What we know is our user are putting lots of small files into the cephFS, now there are around 560 Million files. We didn't see high CPU wait on MDS instance and meta data pool just used around 200MB space.
>> >> >
>> >> > My question is, what is the relationship between the metadata pool and MDS? Is this performance issue caused by the hardware behind meta data pool? Why the meta data pool only used 200MB space, and we saw 3k iops on each of these three ssds, why can't MDS cache all these 200MB into memory?
>> >> >
>> >> > Thanks very much!
>> >> >
>> >> >
>> >> > Best Regards,
>> >> >
>> >> > Albert
>> >> >
>> >> > _______________________________________________
>> >> > ceph-users mailing list
>> >> > ceph-users@xxxxxxxxxxxxxx
>> >> > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com



[Index of Archives]     [Information on CEPH]     [Linux Filesystem Development]     [Ceph Development]     [Ceph Large]     [Ceph Dev]     [Linux USB Development]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [xfs]


  Powered by Linux