Skip to main content

12 posts tagged with "tooling"

View All Tags

· 2 min read
Joachim Kraftmayer

motivation​

we have tested ceph s3 in openstack swift intensively before. We were interested in the behavior of the radosgw stack in ceph. We paid particular attention to the size and number of objects in relation to the resource consumption of the radosgw process. Effects on response latencies of radosgw were also important to us. To be able to plan the right sizing of the physical and virtual environments.

technical topics​​

From a technical point of view, we were interested in the behavior of radosgw in the following topics.

  • dynamic bucket sharding
  • http frontend difference between Civetweb and Beast
  • index pool io pattern and latencies
  • data pool io pattern and latencies with erasure-coded and replicated pools
  • fast_read vs. standard read for workloads with large and small objects.

requirements​

when choosing the right tool, it was important for us to be able to test both small and large ceph clusters with several thousand osds.

We want to use the test results as a file for evaluation as well as have a graphical representation as timeseries data.

For timeseries data we rely on the standard stack with Grafana, Prometheus and Thanos.

the main prometheus exporters we use are ceph-mgr-exporter and node-exporter.

load and performance tools​​

CBT - The Ceph Benchmarking Tool​

CBT is a testing harness written in python

https://github.com/ceph/cbt

s3 - tests​

This is a set of unofficial Amazon AWS S3 compatibility tests

https://github.com/ceph/s3-tests

COSBench - Cloud Object Storage Benchmark​

COSBench is a benchmarking tool to measure the performance of Cloud Object Storage services.

https://github.com/intel-cloud/cosbench

Gosbench​

Gosbench is the Golang reimplementation of Cosbench. It is a distributed S3 performance benchmark tool with Prometheus exporter leveraging the official Golang AWS SDK

https://github.com/mulbc/gosbench

hsbench​

hsbench is an S3 compatable benchmark originally based on wasabi-tech/s3-benchmark.

https://github.com/markhpc/hsbench

Warp​

Minio - S3 benchmarking tool.

https://github.com/minio/warp

the tool of our choice​​

getput​

getput can be run individually on a test client.

gpsuite is responsible for synchronization and scaling across any number of test clients. Communication takes place via ssh keys and the simultaneous start of all s3 test clients is synchronized over a common time base.

Installation on linux as script or as container is supported.

https://github.com/markseger/getput

· One min read
Joachim Kraftmayer

If you don't want to set flags for the whole cluster, like noout or noup. Then you can also use ceph osd set-group and ceph osd unset-group to set the appropriate flag for a group of osds or even whole hosts.

ceph osd set-group <flags> <who>
ceph osd unset-group <flags> <who>

for example set noout for a whole host with osds

ceph osd set-group noout clyso-ceph-node3
``

```bash
root@clyso-ceph-node1:~# ceph health detail
HEALTH_WARN 1 OSDs or CRUSH {nodes, device-classes} have {NOUP,NODOWN,NOIN,NOOUT} flags set
[WRN] OSD_FLAGS: 1 OSDs or CRUSH {nodes, device-classes} have {NOUP,NODOWN,NOIN,NOOUT} flags set
    host clyso-ceph-node3 has flags noout
ceph osd unset-group noout clyso-ceph-node3
root@clyso-ceph-node1:~# ceph health detail
HEALTH_OK
root@clyso-ceph-node1:

Sources:

docs.ceph.com/en/quincy/rados/operations/health-checks/#osd-flags

· 2 min read
Joachim Kraftmayer

ceph-volume can be used to create for a existing OSD a new WAL/DB on a faster device without the need to recreate the OSD.

ceph-volume lvm new-db --osd-id 15 --osd-fsid FSID --target cephdb/cephdb1
--> NameError: name 'get_first_lv' is not defined
this is a bug in ceph-volume v16.2.7 that will be fixed in v16.2.8 
[https://github.com/ceph/ceph/pull/44209](https://github.com/ceph/ceph/pull/44209)

First create a new Logical Volume on the Device that will hold the new WAL/DB

vgcreate cephdb /dev/sdb
Volume group "cephdb" successfully created
lvcreate -L 100G -n cephdb1 cephdb
Logical volume "cephdb1" created.

Now stop running OSD and if it was deactivated ( cephadm ) then activate it on the host

systemctl stop ceph-FSID@osd.0.service
ceph-volume lvm activate --all --no-systemd

Create new WAL/DB on new Device

ceph-volume lvm new-db --osd-id 0 --osd-fsid OSD-FSID --target cephdb/cephdb1
--> Making new volume at /dev/cephdb/cephdb1 for OSD: 0 (/var/lib/ceph/osd/ceph-0)
Running command: /bin/chown -h ceph:ceph /var/lib/ceph/osd/ceph-0/block.db
Running command: /bin/chown -R ceph:ceph /dev/dm-1
--> New volume attached.

Migrate existing WAL/DB to new Device

ceph-volume lvm migrate --osd-id 0 --osd-fsid OSD-FSID --from data --target cephdb/cephdb1
--> Migrate to existing, Source: ['--devs-source', '/var/lib/ceph/osd/ceph-0/block'] Target: /var/lib/ceph/osd/ceph-0/block.db
--> Migration successful.

Deactivate OSD and start it

ceph-volume lvm deactivate 0
Running command: /bin/umount -v /var/lib/ceph/osd/ceph-0
 stderr: umount: /var/lib/ceph/osd/ceph-0 unmounted
systemctl start ceph-FSID@osd.0.service

· One min read
Joachim Kraftmayer

the mysql database was running out of space and we have to increase the pvc size for it.

nothing simple than that. just verified if we have enough free space left.

edited the pvc defintion of the mysql statefulset and set the spec storage size to 20 Gi.

a few seconds later the mysql database space was doubled.

**ceph version 15.2.8

**ceph-csi version: 3.2.1

kubectl edit pvc data-mysql-0 -n mysql

· One min read
Joachim Kraftmayer
  • osd max backfills: This is the maximum number of backfill operations allowed to/from OSD. The higher the number, the quicker the recovery, which might impact overall cluster performance until recovery finishes.
  • osd recovery max active: This is the maximum number of active recover requests. Higher the number, quicker the recovery, which might impact the overall cluster performance until recovery finishes.
  • osd recovery op priority: This is the priority set for recovery operation. Lower the number, higher the recovery priority. Higher recovery priority might cause performance degradation until recovery completes.
ceph tell 'osd.*' injectargs --osd-max-backfills=2 --osd-recovery-max-active=2

Recommendation​

Start in small steps, observe the Ceph status, client IOPs and throughput and then continue to increase in small steps.

In the producton with regard to the applications and hardware infrastructure, we recommend setting these settings back to default as soon as possible.

Sources​

https://www.suse.com/support/kb/doc/?id=000019693

· One min read
Joachim Kraftmayer

Bluestore/RocksDB will only put the next level up size of DB on flash if the whole size will fit. These sizes are roughly 3GB,30GB,300GB. Anything in-between those sizes are pointless. Only ~3GB of SSD will ever be used out of a 28GB partition. Likewise a 240GB partition is also pointless as only ~30GB will be used.

How do I find the right SSD/NVMe partition size for hot DB.

https://github.com/facebook/rocksdb/wiki/Leveled-Compaction

· One min read
Joachim Kraftmayer

When commissioning a cluster, it is always advisable to log and evaluate the ceph osd bench results.

The values can also be helpful for performance analysis in a productive Ceph cluster.

ceph tell osd.<int|*> bench {<int>} {<int>} {<int>}

OSD benchmark: write <count> <size> -byte objects, (default 1G size 4MB)

osd_bench_max_block_size=65536 kB

Example:

1G size 4MB (default)

ceph tell osd.* bench

1G size 64MB

ceph tell osd.* bench 1073741824 67108864