After going through the kernel code, the 'rasize' option seem never making much sense any more after netfs and 'ceph_netfs_expand_readahead()' were introduced to kernel ceph. The ceph client will try to expand the readahead windows twice and the second time it will always align the len down and up to the bounders of the 'stripe_unit'. Anyway this is unfriendly to small size reading case. Thanks - Xiubo On 15/03/2023 02:23, Ashu Pachauri wrote:
Got the answer to my own question; posting here if someone else encounters the same problem. The issue is that the default stripe size in a cephfs mount is 4 MB. If you are doing small reads (like 4k reads in the test I posted) inside the file, you'll end up pulling at least 4MB to the client (and then discarding most of the pulled data) even if you set readahead to zero. So, the solution for us was to set a lower stripe size, which aligns better with our workloads.
Thanks and Regards, Ashu Pachauri
On Fri, Mar 10, 2023 at 9:41 PM Ashu Pachauri <ashu210890@gmail.com> wrote:
Also, I am able to reproduce the network read amplification when I try to do very small reads from larger files. e.g.
for i in $(seq 1 10000); do dd if=test_${i} of=/dev/null bs=5k count=10 done
This piece of code generates a network traffic of 3.3 GB while it actually reads approx 500 MB of data.
Thanks and Regards, Ashu Pachauri
On Fri, Mar 10, 2023 at 9:22 PM Ashu Pachauri <ashu210890@gmail.com> wrote:
We have an internal use case where we back the storage of a proprietary database by a shared file system. We noticed something very odd when testing some workload with a local block device backed file system vs cephfs. We noticed that the amount of network IO done by cephfs is almost double compared to the IO done in case of a local file system backed by an attached block device.
We also noticed that CephFS thrashes through the page cache very quickly compared to the amount of data being read and think that the two issues might be related. So, I wrote a simple test.
1. I wrote 10k files 400KB each using dd (approx 4 GB data). 2. I dropped the page cache completely. 3. I then read these files serially, again using dd. The page cache usage shot up to 39 GB for reading such a small amount of data.
Following is the code used to repro this in bash:
for i in $(seq 1 10000); do dd if=/dev/zero of=test_${i} bs=4k count=100 done
sync; echo 1 > /proc/sys/vm/drop_caches
for i in $(seq 1 10000); do dd if=test_${i} of=/dev/null bs=4k count=100 done
The ceph version being used is: ceph version 15.2.13 (c44bc49e7a57a87d84dfff2a077a2058aa2172e2) octopus (stable)
The ceph configs being overriden: WHO MASK LEVEL OPTION VALUE RO mon advanced auth_allow_insecure_global_id_reclaim false
mgr advanced mgr/balancer/mode upmap
mgr advanced mgr/dashboard/server_addr 127.0.0.1 * mgr advanced mgr/dashboard/server_port 8443 * mgr advanced mgr/dashboard/ssl false * mgr advanced mgr/prometheus/server_addr 0.0.0.0 * mgr advanced mgr/prometheus/server_port 9283 * osd advanced bluestore_compression_algorithm lz4
osd advanced bluestore_compression_mode aggressive osd advanced bluestore_throttle_bytes 536870912 osd advanced osd_max_backfills 3
osd advanced osd_op_num_threads_per_shard_ssd 8 * osd advanced osd_scrub_auto_repair true
mds advanced client_oc false
mds advanced client_readahead_max_bytes 4096
mds advanced client_readahead_max_periods 1
mds advanced client_readahead_min 0
mds basic mds_cache_memory_limit 21474836480 client advanced client_oc false
client advanced client_readahead_max_bytes 4096
client advanced client_readahead_max_periods 1
client advanced client_readahead_min 0
client advanced fuse_disable_pagecache false
The cephfs mount options (note that readahead was disabled for this test): /mnt/cephfs type ceph (rw,relatime,name=cephfs,secret=<hidden>,acl,rasize=0)
Any help or pointers are appreciated; this is a major performance issue for us.
Thanks and Regards, Ashu Pachauri
ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io
-- Best Regards, Xiubo Li (李秀波) Email: xiubli@redhat.com/xiubli@ibm.com Slack: @Xiubo Li