CBT - The Ceph Benchmarking Tool

INTRODUCTION

CBT is a testing harness written in python that can automate a variety of tasks related to testing the performance of Ceph clusters. CBT does not install Ceph packages, it is expected that this will be done prior to utilizing CBT. CBT can create OSDs at the beginning of a test run, optionally recreate OSDs between test runs, or simply run against an existing cluster. CBT records system metrics with collectl, it can optionally collect more information using a number of tools including perf, blktrace, and valgrind. In addition to basic benchmarks, CBT can also do advanced testing that includes automated OSD outages, erasure coded pools, and cache tier configurations. The main benchmark modules are explained below.

radosbench

RADOS bench testing uses the rados binary that comes with the ceph-common package. It contains a benchmarking facility that exercises the cluster by way of librados, the low level native object storage API provided by Ceph. Currently, the RADOS bench module creates a pool for each client.

librbdfio

The librbdfio benchmark module is the simplest way of testing block storage performance of a Ceph cluster. Recent releases of the flexible IO tester (fio) provide a RBD ioengine. This allows fio to test block storage performance of RBD volumes without KVM/QEMU configuration, through the userland librbd libraries. These libraries are the same ones used by the, QEMU backend, so it allows a approximation to KVM/QEMU performance.

kvmrbdfio

The kvmrbdfio benchmark uses the flexible IO tester (fio) to exercise a RBD volume that has been attached to a KVM instance. It requires that the instances be created and have RBD volumes attached prior to using CBT. This module is commonly used to benchmark RBD backed Cinder volumes that have been attached to instances created with OpenStack. Alternatively the instances could be provisioned using something along the lines of Vagrant or Virtual Machine Manager.

rbdfio

The rbdfio benchmark uses the flexible IO tester (fio) to excercise a RBD volume that has been mapped to a block device using the KRBD kernel driver. This module is most relevant for simulating the data path for applications that need a block device, but wont for whatever reason be ran inside a virtual machine.

PREREQUISITES

CBT uses several libraries and tools to run:

python3-yaml - A YAML library for python used for reading configuration files.
python3-lxml - Powerful and Pythonic XML processing library combining libxml2/libxslt with the ElementTree API
ssh (and scp) - remote secure command executation and data transfer
pdsh (and pdcp) - a parallel ssh and scp implementation
ceph - A scalable distributed storage system

Note that pdsh is not packaged for RHEL7 and CentOS 7 based distributations at this time, though the rawhide pdsh packages install and are usable. The RPMs for these packages are available here:

ftp://rpmfind.net/linux/fedora/linux/releases/23/Everything/x86_64/os/Packages/p/pdsh-2.31-4.fc23.x86_64.rpm
ftp://rpmfind.net/linux/fedora/linux/releases/23/Everything/x86_64/os/Packages/p/pdsh-rcmd-rsh-2.31-4.fc23.x86_64.rpm
ftp://rpmfind.net/linux/fedora/linux/releases/23/Everything/x86_64/os/Packages/p/pdsh-rcmd-ssh-2.31-4.fc23.x86_64.rpm

Optional tools and benchmarks can be used if desired:

collectl - system data collection
blktrace - block device io tracing
seekwatcher - create graphs and movies from blktrace data
perf - system and process profiling
valgrind - runtime memory and cpu profiling of specific processes
fio - benchmark suite with integrated posix, libaio, and librbd support
cosbench - object storage benchmark from Intel

USER AND NODE SETUP

In addition to the above software, a number of nodes must be available to run tests. These are divided into several categories. Multiple categories can contain the same host if it is assuming multiple roles (running OSDs and a mon for instance).

head - node where general ceph commands are run
clients - nodes that will run benchmarks or other client tools
osds - nodes where OSDs will live
rgws - nodes where rgw servers will live
mons - nodes where mons will live

A user may also be specified to run all remote commands. The host that is used to run cbt must be able to issue passwordless ssh commands as the specified user. This can be accomplished by creating a passwordless ssh key:

ssh-keygen -t dsa

and copying the resulting public key in the ~/.ssh to the ~/.ssh/authorized_key file on all remote hosts.

This user must also be able to run certain commands with sudo. The easiest method to enable this is to simply enable blanket passwordless sudo access for this user, though this is only appropriate in laboratory environments. This may be acommplished by running visudo and adding something like:

# passwordless sudo for cbt
<user>    ALL=(ALL)       NOPASSWD:ALL

Where <user> is the user that will have password sudo access.
Please see your OS documentation for specific details.

In addition to the above, it will be required to add all osds and mons into the list of known hosts for ssh in order to perform properly. Otherwise, the benchmarking tests will not be able to run.

Note that the pdsh command could have difficulties if the sudoers file requires tty. If this is the case, commend out the Defaults requiretty line in visudo.

DISK PARTITIONING

Currently CBT looks for specific partition labels in /dev/disk/by-partlabel for the Ceph OSD data and journal partitions.
At some point in the future this will be made more flexible, for now this is the expected behavior. Specifically on each OSD host partitions should be specified with the following gpt labels:

osd-device-<num>-data
osd-device-<num>-journal

where <num> is a device ordered starting at 0 and ending with the last device on the system. Currently cbt assumes that all nodes in the system have the same number of devices. A script is available that shows an example of how we create partition labels in our test lab here:

https://github.com/ceph/cbt/blob/master/tools/mkpartmagna.sh

CREATING A YAML FILE

CBT yaml files have a basic structure where you define a cluster and a set of benchmarks to run against it. For example, the following yaml file creates a single node cluster on a node with hostname "burnupiX". A pool profile is defined for a 1x replication pool using 256 PGs, and that pool is used to run RBD performance tests using fio with the librbd engine.

cluster:
  user: 'nhm'
  head: "burnupiX"
  clients: ["burnupiX"]
  osds: ["burnupiX"]
  mons:
    burnupiX:
      a: "127.0.0.1:6789"
  osds_per_node: 1
  fs: 'xfs'
  mkfs_opts: '-f -i size=2048'
  mount_opts: '-o inode64,noatime,logbsize=256k'
  conf_file: '/home/nhm/src/ceph-tools/cbt/newstore/ceph.conf.1osd'
  iterations: 1
  use_existing: False
  clusterid: "ceph"
  tmp_dir: "/tmp/cbt"
  pool_profiles:
    rbd:
      pg_size: 256
      pgp_size: 256
      replication: 1
benchmarks:
  librbdfio:
    time: 300
    vol_size: 16384
    mode: [read, write, randread, randwrite]
    op_size: [4194304, 2097152, 1048576]
    concurrent_procs: [1]
    iodepth: [64]
    osd_ra: [4096]
    cmd_path: '/home/nhm/src/fio/fio'
    pool_profile: 'rbd'

An associated ceph.conf.1osd file is also defined with various settings that are to be used in this test:

[global]
        osd pool default size = 1
        auth cluster required = none
        auth service required = none
        auth client required = none
        keyring = /tmp/cbt/ceph/keyring
        osd pg bits = 8  
        osd pgp bits = 8
        log to syslog = false
        log file = /tmp/cbt/ceph/log/$name.log
        public network = 192.168.10.0/24
        cluster network = 192.168.10.0/24
        rbd cache = true
        osd scrub load threshold = 0.01
        osd scrub min interval = 137438953472
        osd scrub max interval = 137438953472
        osd deep scrub interval = 137438953472
        osd max scrubs = 16
        filestore merge threshold = 40
        filestore split multiple = 8
        osd op threads = 8
        mon pg warn max object skew = 100000
        mon pg warn min per osd = 0
        mon pg warn max per osd = 32768

[mon]
        mon data = /tmp/cbt/ceph/mon.$id
        
[mon.a]
        host = burnupiX 
        mon addr = 127.0.0.1:6789

[osd.0]
        host = burnupiX
        osd data = /tmp/cbt/mnt/osd-device-0-data
        osd journal = /dev/disk/by-partlabel/osd-device-0-journal

To run this benchmark suite, cbt is launched with an output archive directory to store the results and the yaml configuration file to use:

cbt.py --archive=<archive dir> ./mytests.yaml

You can also specify the ceph.conf file to use by specifying it on the commandline:

cbt.py --archive=<archive dir> --conf=./ceph.conf.1osd ./mytests.yaml

In this way you can mix and match ceph.conf files and yaml test configuration files to create parametric sweeps of tests. A script in the tools directory called mkcephconf.py lets you automatically generate hundreds or thousands of ceph.conf files from defined ranges of different options that can then be used with cbt in this way.

CONCLUSION

There are many additional and powerful ways you can use cbt that are not yet covered in this document. As time goes on we will try to provide better examples and documentation for these features. For now, it's best to look at the examples, look at the code, and ask questions!

bkukanov/cbt