Re: Multipart uploads with partsizes larger than 16MiB failing on Nautilus

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

 



On Thu, Sep 10, 2020 at 10:19 AM shubjero <shubjero@xxxxxxxxx> wrote:
>
> Hi Casey,
>
> I was never setting rgw_max_chunk_size in my ceph.conf so it must have
> been default? Funny enough I dont even see this configuration
> parameter in the documentation
> https://docs.ceph.com/docs/nautilus/radosgw/config-ref/ .
>
> Armed with your information I tried setting the following in my ceph.conf:
>
> root@ceph-1:~# ceph --admin-daemon
> /var/run/ceph/ceph-client.rgw.ceph-1.28726.94406486979736.asok config
> show | egrep "rgw_max_chunk_size|rgw_put_obj_min|rgw_obj_stripe_size"
>     "rgw_max_chunk_size": "67108864",
>     "rgw_obj_stripe_size": "67108864",
>     "rgw_put_obj_min_window_size": "67108864",
>
> And with this configuration I was able to upload with large part sizes
> (2GB) using the aws client without error.

are you sure there's a benefit to using such large part sizes? a
smaller part size should allow the client to stream more uploads at a
time. it also makes recovery much cheaper; if a 2GB PUT request times
out, the client will retry and send the entire 2GB again. with a
smaller part size, the server can commit this data more frequently and
limit the amount of bandwidth wasted on retries

>
> Do you know if there is any expected performance improvement with
> larger chunk/stripe/window sizes? As I said previously our use case is
> dealing with very large genomic files being uploaded and downloaded
> (average is probably 100GB per file).

rgw_max_chunk_size specifies how much data we'll send in a single osd
request. rgw_obj_stripe_size specifies how much data we'll write to a
single rados object before creating a new stripe/object.
rgw_put_obj_min_window_size specifies how much object data we'll
buffer in memory as we stream chunks out to their osds

i don't think we saw any benefit from chunk sizes over 4M, but you're
welcome to experiment and measure that in your environment. generally
you want a rgw_obj_stripe_size == rgw_max_chunk_size so that each of
your writes go to a different rados object; if, for example, your
stripe size was 2x the chunk size, we would write two chunks to each
rados object - but the osd has to apply these writes sequentially, so
you lose some parallelism this way

regarding rgw_put_obj_min_window_size, the number of parallel writes
we can do is equal to (rgw_put_obj_min_window_size /
rgw_max_chunk_size). in a default configuration, this is 16M/4M = 4.
you can experiment with a larger multiplier here, but do take overall
memory usage into account! If rgw_max_concurrent_requests is 1024 and
all of those are large PUT requests, then we'd use up to
(rgw_max_concurrent_requests * rgw_put_obj_min_window_size) or 16G of
memory

in general, i think the default tunings should perform well here. if
you have a lot of memory to work with on rgw nodes, you can experiment
with larger values of rgw_put_obj_min_window_size

>
> On Wed, Sep 9, 2020 at 11:29 AM Casey Bodley <cbodley@xxxxxxxxxx> wrote:
> >
> > What is your rgw_max_chunk_size? It looks like you'll get these
> > EDEADLK errors when rgw_max_chunk_size > rgw_put_obj_min_window_size,
> > because we try to write in units of chunk size but the window is too
> > small to write a single chunk.
> >
> > On Wed, Sep 9, 2020 at 8:51 AM shubjero <shubjero@xxxxxxxxx> wrote:
> > >
> > > Will do Matt
> > >
> > > On Tue, Sep 8, 2020 at 5:36 PM Matt Benjamin <mbenjami@xxxxxxxxxx> wrote:
> > > >
> > > > thanks, Shubjero
> > > >
> > > > Would you consider creating a ceph tracker issue for this?
> > > >
> > > > regards,
> > > >
> > > > Matt
> > > >
> > > > On Tue, Sep 8, 2020 at 4:13 PM shubjero <shubjero@xxxxxxxxx> wrote:
> > > > >
> > > > > I had been looking into this issue all day and during testing found
> > > > > that a specific configuration option we had been setting for years was
> > > > > the culprit. Not setting this value and letting it fall back to the
> > > > > default seems to have fixed our issue with multipart uploads.
> > > > >
> > > > > If you are curious, the configuration option is rgw_obj_stripe_size
> > > > > which was being set to 67108864 bytes (64MiB). The default is 4194304
> > > > > bytes (4MiB). This is a documented option
> > > > > (https://docs.ceph.com/docs/nautilus/radosgw/config-ref/) and from my
> > > > > testing it seems like using anything but the default (only tried
> > > > > larger values) breaks multipart uploads.
> > > > >
> > > > > On Tue, Sep 8, 2020 at 12:12 PM shubjero <shubjero@xxxxxxxxx> wrote:
> > > > > >
> > > > > > Hey all,
> > > > > >
> > > > > > I'm creating a new post for this issue as we've narrowed the problem
> > > > > > down to a partsize limitation on multipart upload. We have discovered
> > > > > > that in our production Nautilus (14.2.11) cluster and our lab Nautilus
> > > > > > (14.2.10) cluster that multipart uploads with a configured part size
> > > > > > of greater than 16777216 bytes (16MiB) will return a status 500 /
> > > > > > internal server error from radosgw.
> > > > > >
> > > > > > So far I have increased the following rgw settings/values that looked
> > > > > > suspect, without any success/improvement with partsizes.
> > > > > > Such as:
> > > > > >     "rgw_get_obj_window_size": "16777216",
> > > > > >     "rgw_put_obj_min_window_size": "16777216",
> > > > > >
> > > > > > I am trying to determine if this is because of a conservative default
> > > > > > setting somewhere that I don't know about or if this is perhaps a bug?
> > > > > >
> > > > > > I would appreciate it if someone on Nautilus with rgw could also test
> > > > > > / provide feedback. It's very easy to reproduce and configuring your
> > > > > > partsize with aws2cli requires you to put the following in your aws
> > > > > > 'config'
> > > > > > s3 =
> > > > > >   multipart_chunksize = 32MB
> > > > > >
> > > > > > rgw server logs during a failed multipart upload (32MB chunk/partsize):
> > > > > > 2020-09-08 15:59:36.054 7f2d32fa6700  1 ====== starting new request
> > > > > > req=0x55953dc36930 =====
> > > > > > 2020-09-08 15:59:36.082 7f2d32fa6700 -1 res_query() failed
> > > > > > 2020-09-08 15:59:36.138 7f2d32fa6700  1 ====== req done
> > > > > > req=0x55953dc36930 op status=0 http_status=200 latency=0.0839988s
> > > > > > ======
> > > > > > 2020-09-08 16:00:07.285 7f2d3dfbc700  1 ====== starting new request
> > > > > > req=0x55953dc36930 =====
> > > > > > 2020-09-08 16:00:07.285 7f2d3dfbc700 -1 res_query() failed
> > > > > > 2020-09-08 16:00:07.353 7f2d00741700  1 ====== starting new request
> > > > > > req=0x55954dd5e930 =====
> > > > > > 2020-09-08 16:00:07.357 7f2d00741700 -1 res_query() failed
> > > > > > 2020-09-08 16:00:07.413 7f2cc56cb700  1 ====== starting new request
> > > > > > req=0x55953dc02930 =====
> > > > > > 2020-09-08 16:00:07.417 7f2cc56cb700 -1 res_query() failed
> > > > > > 2020-09-08 16:00:07.473 7f2cb26a5700  1 ====== starting new request
> > > > > > req=0x5595426f6930 =====
> > > > > > 2020-09-08 16:00:07.473 7f2cb26a5700 -1 res_query() failed
> > > > > > 2020-09-08 16:00:09.465 7f2d3dfbc700  0 WARNING: set_req_state_err
> > > > > > err_no=35 resorting to 500
> > > > > > 2020-09-08 16:00:09.465 7f2d3dfbc700  1 ====== req done
> > > > > > req=0x55953dc36930 op status=-35 http_status=500 latency=2.17997s
> > > > > > ======
> > > > > > 2020-09-08 16:00:09.549 7f2d00741700  0 WARNING: set_req_state_err
> > > > > > err_no=35 resorting to 500
> > > > > > 2020-09-08 16:00:09.549 7f2d00741700  1 ====== req done
> > > > > > req=0x55954dd5e930 op status=-35 http_status=500 latency=2.19597s
> > > > > > ======
> > > > > > 2020-09-08 16:00:09.605 7f2cc56cb700  0 WARNING: set_req_state_err
> > > > > > err_no=35 resorting to 500
> > > > > > 2020-09-08 16:00:09.609 7f2cc56cb700  1 ====== req done
> > > > > > req=0x55953dc02930 op status=-35 http_status=500 latency=2.19597s
> > > > > > ======
> > > > > > 2020-09-08 16:00:09.641 7f2cb26a5700  0 WARNING: set_req_state_err
> > > > > > err_no=35 resorting to 500
> > > > > > 2020-09-08 16:00:09.641 7f2cb26a5700  1 ====== req done
> > > > > > req=0x5595426f6930 op status=-35 http_status=500 latency=2.16797s
> > > > > > ======
> > > > > >
> > > > > > awscli client side output during a failed multipart upload:
> > > > > > root@jump:~# aws --no-verify-ssl --endpoint-url
> > > > > > http://lab-object.cancercollaboratory.org:7480 s3 cp 4GBfile
> > > > > > s3://troubleshooting
> > > > > > upload failed: ./4GBfile to s3://troubleshooting/4GBfile An error
> > > > > > occurred (UnknownError) when calling the UploadPart operation (reached
> > > > > > max retries: 2): Unknown
> > > > > >
> > > > > > Thanks,
> > > > > >
> > > > > > Jared Baker
> > > > > > Cloud Architect for the Cancer Genome Collaboratory
> > > > > > Ontario Institute for Cancer Research
> > > > > _______________________________________________
> > > > > ceph-users mailing list -- ceph-users@xxxxxxx
> > > > > To unsubscribe send an email to ceph-users-leave@xxxxxxx
> > > > >
> > > >
> > > >
> > > > --
> > > >
> > > > Matt Benjamin
> > > > Red Hat, Inc.
> > > > 315 West Huron Street, Suite 140A
> > > > Ann Arbor, Michigan 48103
> > > >
> > > > http://www.redhat.com/en/technologies/storage
> > > >
> > > > tel.  734-821-5101
> > > > fax.  734-769-8938
> > > > cel.  734-216-5309
> > > >
> > > _______________________________________________
> > > 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



[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