Contents
SwiftStack Benchmark Suite (ssbench) is a flexible and scalable
benchmarking tool for the OpenStack Swift object storage system.
The ssbench-master run-scenario command will run benchmark "scenarios"
against an
OpenStack Swift cluster, utilizing one or more distributed ssbench-worker
processes, saving statistics about the run to a file. The ssbench-master
report-scenario command can then generate a
report from the saved statstics. By default, ssbench-master run-scenario
will generate a report to STDOUT immediately following a benchmark run in
addition to saving the raw results to a file.
Coordination between the ssbench-master and one or more ssbench-worker
processes is managed through a pair of PyZMQ sockets. This
allows ssbench-master to distribute the benchmark run across many, many
client servers while still coordinating the entire run (each worker can be
given a job referencing an object created by a different worker).
Continuous Integration:
Code Quality:
I apologize for this stupid dependency dance with Ubuntu (tested with 12.04
LTS Precise). With the --noop benchmark, gevent-zeromq is about 25%
faster than pyzmq 2.2.0.1's zmq.green module, so I consider the annoying
gevent-zeromq dependency worth it. The gevent-zeormq
Cython build doesn't work with Ubuntu 12.04's Python's distribute, and Cython
has to be installed in a prior "pip" command to be recognized by
gevent-zeromq's setup.py:
$ sudo apt-get install -y python-dev python-pip 'g++' libzmq-dev libevent-dev $ sudo pip install --upgrade distribute $ sudo pip install Cython gevent pyzmq==2.2.0 $ sudo pip install ssbench
Installation on Fedora 18 using its stock Python 2.7:
$ sudo yum install -y gcc gcc-c++ python-setuptools python-devel libevent-devel python-pip zeromq3-devel python-argparse Cython gevent $ sudo pip install Distribute $ sudo pip install --upgrade pyzmq==2.2.0 $ sudo pip install gevent-zeromq $ sudo pip install ssbench
Installation on RHEL 6.4 using its stock Python 2.6:
$ sudo rpm -Uvh http://mirror.pnl.gov/epel/6/i386/epel-release-6-8.noarch.rpm $ sudo yum install -y gcc gcc-c++ python-setuptools python-devel libevent-devel python-pip zeromq3-devel python-argparse Cython $ sudo pip install Distribute $ sudo pip install --upgrade gevent pyzmq==2.2.0 $ sudo pip install gevent-zeromq $ sudo pip install ssbench
Installation on CentOS 6.3 using its stock Python 2.6:
$ sudo rpm -Uvh http://mirror.pnl.gov/epel/6/i386/epel-release-6-8.noarch.rpm $ sudo yum install -y gcc gcc-c++ python-setuptools python-devel libevent-devel python-pip zeromq3-devel $ sudo pip-python install --upgrade argparse distribute Cython gevent pyzmq==2.2.0 $ sudo pip-python install gevent-zeromq $ sudo pip-python install ssbench
On the Mac, I recommend installing Homebrew and using that to install Python 2.7, libevent, and zeromq. I haven't tested a fresh install in a while, but I had far less problems with Cython and gevent-zeormq on OS X, probably because the Homebrew Python was newer than Ubuntu 12.04's?
Then you should be able to just pip install ssbench.
I have not tested ssbench against
gevent 1.0rc2, but according to an old gevent blog post, gevent v1.x will
bundle libev and not require the installation of libevent or
libev_. If you try ssbench with gevent 1.0rc2, please let me know if
and how that works...
A "scenario" (sometimes called a "CRUD scenario") is a utf8-encoded JSON file defining a benchmark run. Specifically, it defines:
- A
namefor the scenario (an arbitrary string) - A
sizeslist of "object size" classes. Each object size class has aname, asize_minminimum object size, asize_maxmaximum object size (in bytes), and an optionalcrud_profilefor just this size. Ifcrud_profileis not given for a size, the top-levelcrud_profilewill be used. Thecrud_profilehere is just like the top-level one, an array of 4 numbers whose relative sizes determine the percent chance of a Create, Read, Update, or Delete operation. Objects created or updated within an object size class will have a size (in bytes) chosen at random uniformly between the minimum and maximum sizes. - An
initial_filesdictionary of initial file-counts per size class. Each size class can have zero or more objects uploaded prior to the benchmark run itself. The proportion of initial files also defines the probability distribution of object sizes during the benchmark run itself. So if a particular object size class is not included ininitial_filesor has a value of 0 ininitial_files, then no objects in that size class will be used during the benchmark run. Each initial object's name and container is deterministic and, as an optimization, if an object of the right name is in the right container, it will not be uploaded again; note that initial objects are not deleted after each benchmark run, so this can speed up subsequent runs quite a bit. - An
operation_countof operations to perform during the benchmark run. An operation is either a CREATE, READ, UPDATE, or DELETE of an object. This value may be overridden for any given run with the-o COUNTflag tossbench-master run-scenario. - A
run_secondsnumber of seconds the benchmark scenario should run. This is mutually exclusive withoperation_count, so only one of those two should be specified. Both values may be overridden with command-line arguments tossbench-master. - A
crud_profilewhich determines the distribution of each kind of operation. For instance,[3, 4, 2, 2]would mean 27% CREATE, 36% READ, 18% UPDATE, and 18% DELETE. - A
user_countwhich determines the maxiumum client concurrency during the benchmark run. The user is responsible for ensuring there are enough workers running to support the scenario's defineduser_count. (Eachssbench-workerprocess uses gevent to achive very efficient concurrency for the benchmark client requests.) This value may be overridden for any given run with the-u COUNTflag tossbench-master run-scenario. - A
container_basewhich is a string used to construct the names of containers used by ssbench. It defaults tossbench, resulting in container names likessbench_000061. - A
container_countwhich determines how many Swift containers are used for the benchmark run. This key is optional in the scenario file and defaults to 100. This value may be overridden for any given run with the-c COUNTflag tossbench-master run-scenario. - A
container_concurrencyvalue which determines the level of client concurrency used byssbench-masterto create the benchmark containers. This value is optional and defaults to 10. - A
delete_aftervalue appends expiring time(in seconds) to all objects. It emulates continuous loads of PUT operation (CREATE and UPDATE) with X-Delete-After header. If setting 0 (or None by default), this feature is disable and all objects will not be expired. This value may be overridden for any given run with the--delete-after DELETE_AFTERflag tossbench-master run-scenario.
For each operation of the benchmark run, a size category is first chosen based
on the relative counts for each size category in the initial_files
dictionary. This probability for each size category appears under the "% Ops"
column in the report. Then an operation type is chosen based on that size
category's CRUD profile (which can be individually specified or may be
inherited from the "top level" CRUD profile).
If each size category has its own CRUD profile, then the overall CRUD profile of the benchmark run will be a weighted average between the values in the "% Ops" column and the CRUD profile of each size category. This weighted average CRUD profile is included in the report on the "CRUD weighted average" line.
ssbench comes with a few canned scenarios, but users are encouraged to
experiment and define their own.
Here is an example JSON scenario file:
{
"name": "Small test scenario",
"sizes": [{
"name": "tiny",
"size_min": 4096,
"size_max": 65536
}, {
"name": "small",
"size_min": 100000,
"size_max": 200000
}],
"initial_files": {
"tiny": 100,
"small": 10
},
"operation_count": 500,
"crud_profile": [3, 4, 2, 2],
"user_count": 7
}
Beware: hand-editing JSON is error-prone. Watch out for trailing commas, in particular.
The ssbench-worker script's usage message may be generated with:
$ ssbench-worker -h
usage: ssbench-worker [-h] [--zmq-host ZMQ_HOST]
[--zmq-work-port ZMQ_WORK_PORT]
[--zmq-results-port ZMQ_RESULTS_PORT] [-c CONCURRENCY]
[--retries RETRIES] [--batch-size COUNT] [-p COUNT] [-v]
worker_id
...
The ssbench-master command requires one sub-command, which is currently
either run-scenario to actually run a benchmark scenario,
report-scenario to report on an existing scenario result data file, or
kill-workers to tell connected ssbench-worker processes not started
with --workers to kill themselves:
usage: ssbench-master [-h] [-v] [-q]
{report-scenario,kill-workers,run-scenario,cleanup-containers}
...
SwiftStack Benchmark (ssbench) version 0.2.20
positional arguments:
{report-scenario,kill-workers,run-scenario,cleanup-containers}
kill-workers Tell all workers to exit.
run-scenario Run CRUD scenario, saving statistics. You must supply
a valid set of v1.0 or v2.0 auth credentials. See
usage message for run-scenario for more details.
report-scenario Generate a report from saved scenario statistics.
Various types of reports may be generated, with the
default being a "textual summary".
cleanup-containers Recursively delete all ssbench containers and their
objects.
optional arguments:
-h, --help show this help message and exit
-v, --verbose Enable more verbose output. (default: False)
-q, --quiet Suppress most output (including progress characters
during run). (default: False)
The run-scenario sub-command of ssbench-master actually
runs a benchmark scenario:
$ ssbench-master run-scenario -h
usage: ssbench-master run-scenario [-h] -f SCENARIO_FILE
[--zmq-bind-ip BIND_IP]
[--zmq-work-port PORT]
[--zmq-results_port PORT] [-V AUTH_VERSION]
[-A AUTH_URL] [-U USER] [-K KEY]
[--os-username <auth-user-name>]
[--os-password <auth-password>]
[--os-tenant-id <auth-tenant-id>]
[--os-tenant-name <auth-tenant-name>]
[--os-auth-url <auth-url>]
[--os-auth-token <auth-token>]
[--os-storage-url <storage-url>]
[--os-region-name <region-name>]
[--os-service-type <service-type>]
[--os-endpoint-type <endpoint-type>]
[--os-cacert <ca-certificate>] [--insecure]
[-S STORAGE_URL] [-T TOKEN] [-c COUNT]
[-u COUNT] [-o COUNT] [-r SECONDS]
[-b BYTES] [--workers COUNT]
[--batch-size COUNT] [--profile] [--noop]
[-k] [--connect-timeout CONNECT_TIMEOUT]
[--network-timeout NETWORK_TIMEOUT]
[-s STATS_FILE] [-R] [--csv]
[--pctile PERCENTILE]
[--delete-after DELETE_AFTER]
...
The report-scenario sub-command of ssbench-master reports on a
previously-run benchmark scenario:
$ ssbench-master report-scenario -h
usage: ssbench-master report-scenario [-h] -s STATS_FILE [-f REPORT_FILE]
[--pctile PERCENTILE] [--csv]
[-r RPS_HISTOGRAM] [--profile]
...
The kill-workers sub-command of ssbench-master kills all
ssbench-worker processes which are pointed at the ssbench-master
ZMQ sockets (this is useful for multi-server benchmark runs where the workers
were not started with ssbench-master's --workers option):
$ ssbench-master kill-workers -h
usage: ssbench-master kill-workers [-h] [--zmq-bind-ip BIND_IP]
[--zmq-work-port PORT]
[--zmq-results_port PORT]
...
The cleanup-containers sub-command of ssbench-master recursively
deletes all ssbench-created containers and objects. It takes all the same
authorization-related options as run-scenario:
$ ssbench-master cleanup-containers -h
usage: ssbench-master cleanup-containers [-h] [-b CONTAINER_BASE]
[-c CONCURRENCY] [-V AUTH_VERSION]
[-A AUTH_URL] [-U USER] [-K KEY]
[--os-username <auth-user-name>]
[--os-password <auth-password>]
[--os-tenant-id <auth-tenant-id>]
[--os-tenant-name <auth-tenant-name>]
[--os-auth-url <auth-url>]
[--os-auth-token <auth-token>]
[--os-storage-url <storage-url>]
[--os-region-name <region-name>]
[--os-service-type <service-type>]
[--os-endpoint-type <endpoint-type>]
[--os-cacert <ca-certificate>]
[--insecure] [-S STORAGE_URL]
[-T TOKEN]
...
ssbench-master supports all the same authentication arguments, with similar
semantics, as python-swiftclient's command-line tool, swift.
For v1.0 authentication, you just need ST_AUTH, ST_USER, and ST_KEY
defined in the environment or overridden/set on the command-line with -A,
-U, and -K, respectively.
For v2.0 authentication (Keystone), it's more complicated and you should refer to Keystone and/or python-swiftclient documentation for more help.
Regardless of which version of authentication is used, you may specify -S
<storage_url> on the command-line to override the Storage URL returned from
the authentication system.
You can bypass your normal load-balancing scheme by telling ssbench-master
to distribute load across a specified set of Storage URLs. This is done by
specifiying one or more -S STORAGE_URL options to ssbench-master. Any
storage URL returned from the auth server will be ignored and a randomly chosen
command-line-specified storage URL will be used instead.
Note that each ssbench-worker process will create a fully-populated
connection pool for each unique -S argument specified. Each connection
pool will contain a number of sockets equal to the -c option (which defaults
to 64). So a large number of unique -S arguments for ssbench-worker
and a large -c value for ssbench-worker processes will not mix well.
Start one or more ssbench-worker processes on each server (each
ssbench-worker process defaults to a maximum gevent-based concurrency
of 64, but the -c option can override that default). Use the
--zmq-host command-line parameter to specify the host on which you will run
ssbench-master.:
bench-host-01$ ssbench-worker -c 1000 --zmq-host bench-host-01 1 & bench-host-01$ ssbench-worker -c 1000 --zmq-host bench-host-01 2 & bench-host-02$ ssbench-worker -c 1000 --zmq-host bench-host-01 3 & bench-host-02$ ssbench-worker -c 1000 --zmq-host bench-host-01 4 &
Finally, run one ssbench-master process which will manage and coordinate
the multi-server benchmark run:
bench-host-01$ ssbench-master run-scenario -f scenarios/very_small.scenario -u 2000 -o 40000
The above example would involve a total client concurrency of 2000, spread
evenly among the four workers on two hosts (bench-host-01 and
bench-host-02). The four workers, as started in the above example,
could support a maximum total client concurrency (-u option to
ssbench-master) up to 4000.
If you only need workers running on the local host, you can do so with a single
command. Simply use the --workers COUNT option to ssbench-master:
$ ssbench-master run-scenario -f scenarios/very_small.scenario -u 4 -c 80 -o 613 --pctile 50 --workers 2
INFO:SwiftStack Benchmark (ssbench version 0.2.14)
INFO:Spawning local ssbench-worker (logging to /tmp/ssbench-worker-local-0.log) with ssbench-worker ... --concurrency 2 --batch-size 1 0
INFO:Spawning local ssbench-worker (logging to /tmp/ssbench-worker-local-1.log) with ssbench-worker ... --concurrency 2 --batch-size 1 1
INFO:Starting scenario run for "Small test scenario"
INFO:Ensuring 80 containers (ssbench_*) exist; concurrency=10...
INFO:Initializing cluster with stock data (up to 4 concurrent workers)
INFO:Starting benchmark run (up to 4 concurrent workers)
Benchmark Run:
X work job raised an exception
. < 1s first-byte-latency
o < 3s first-byte-latency
O < 10s first-byte-latency
* >= 10s first-byte-latency
_ < 1s last-byte-latency (CREATE or UPDATE)
| < 3s last-byte-latency (CREATE or UPDATE)
^ < 10s last-byte-latency (CREATE or UPDATE)
@ >= 10s last-byte-latency (CREATE or UPDATE)
....._........_.._......_.._..__.._.._..._...__...__._..._._..................
....._.._....__........._.._._......__.._.._._......._..__.._....._..._...__._
...._......_....____....__._.........._...._...._......._....__._.._._..__._..
....__.._..._._._....._......_...._...__...._...___.........._.._._..___..._._
....._._....__.............._.__..._...._...._...._._.._....___........_.__.._
_..__._.__.._.................__......._......._...._.____...._.._....._...._.
..._.............__.._..._.._.._._._._...._.._.._....__._._........_......_.__
.........._._...._.._.........._........_._.._....._......._....._.
INFO:Deleting population objects from cluster
INFO:Calculating statistics...
Small test scenario (generated with ssbench version 0.2.14)
Worker count: 2 Concurrency: 4 Ran 2013-06-07 17:23:16 UTC to 2013-06-07 17:23:22 UTC (5s)
Object expiration (X-Delete-After): None (sec)
% Ops C R U D Size Range Size Name
91% % 10 75 15 0 4 kB - 8 kB tiny
9% % 10 75 15 0 20 kB - 40 kB small
---------------------------------------------------------------------
10 75 15 0 CRUD weighted average
TOTAL
Count: 613 ( 0 error; 0 retries: 0.00%) Average requests per second: 118.7
min max avg std_dev 50%-ile Worst latency TX ID
First-byte latency: 0.004 - 0.044 0.017 ( 0.008) 0.016 (all obj sizes) txe026893bbf09486c83fcdb629f6f25a3
Last-byte latency: 0.004 - 0.157 0.029 ( 0.024) 0.019 (all obj sizes) tx6f988120ec5044329f817-0051b21708
First-byte latency: 0.004 - 0.044 0.016 ( 0.007) 0.016 ( tiny objs) tx1d35c8e273bf4bbeb6298-0051b21705
Last-byte latency: 0.004 - 0.157 0.028 ( 0.024) 0.019 ( tiny objs) tx6f988120ec5044329f817-0051b21708
First-byte latency: 0.005 - 0.044 0.018 ( 0.008) 0.016 ( small objs) txe026893bbf09486c83fcdb629f6f25a3
Last-byte latency: 0.005 - 0.120 0.031 ( 0.026) 0.021 ( small objs) tx87bf30db5a70412b97a5c71ae60036c1
CREATE
Count: 64 ( 0 error; 0 retries: 0.00%) Average requests per second: 12.5
min max avg std_dev 50%-ile Worst latency TX ID
First-byte latency: N/A - N/A N/A ( N/A ) N/A (all obj sizes)
Last-byte latency: 0.024 - 0.157 0.067 ( 0.023) 0.060 (all obj sizes) tx6f988120ec5044329f817-0051b21708
First-byte latency: N/A - N/A N/A ( N/A ) N/A ( tiny objs)
Last-byte latency: 0.024 - 0.157 0.064 ( 0.022) 0.059 ( tiny objs) tx6f988120ec5044329f817-0051b21708
First-byte latency: N/A - N/A N/A ( N/A ) N/A ( small objs)
Last-byte latency: 0.061 - 0.120 0.087 ( 0.020) 0.089 ( small objs) tx87bf30db5a70412b97a5c71ae60036c1
READ
Count: 459 ( 0 error; 0 retries: 0.00%) Average requests per second: 88.9
min max avg std_dev 50%-ile Worst latency TX ID
First-byte latency: 0.004 - 0.044 0.017 ( 0.008) 0.016 (all obj sizes) txe026893bbf09486c83fcdb629f6f25a3
Last-byte latency: 0.004 - 0.044 0.017 ( 0.008) 0.016 (all obj sizes) txe026893bbf09486c83fcdb629f6f25a3
First-byte latency: 0.004 - 0.044 0.016 ( 0.007) 0.016 ( tiny objs) tx1d35c8e273bf4bbeb6298-0051b21705
Last-byte latency: 0.004 - 0.044 0.017 ( 0.007) 0.016 ( tiny objs) tx1d35c8e273bf4bbeb6298-0051b21705
First-byte latency: 0.005 - 0.044 0.018 ( 0.008) 0.016 ( small objs) txe026893bbf09486c83fcdb629f6f25a3
Last-byte latency: 0.005 - 0.044 0.019 ( 0.008) 0.017 ( small objs) txe026893bbf09486c83fcdb629f6f25a3
UPDATE
Count: 90 ( 0 error; 0 retries: 0.00%) Average requests per second: 18.1
min max avg std_dev 50%-ile Worst latency TX ID
First-byte latency: N/A - N/A N/A ( N/A ) N/A (all obj sizes)
Last-byte latency: 0.021 - 0.143 0.062 ( 0.021) 0.061 (all obj sizes) tx9a502107a0c246e69a987d120a2b9919
First-byte latency: N/A - N/A N/A ( N/A ) N/A ( tiny objs)
Last-byte latency: 0.021 - 0.143 0.062 ( 0.022) 0.061 ( tiny objs) tx9a502107a0c246e69a987d120a2b9919
First-byte latency: N/A - N/A N/A ( N/A ) N/A ( small objs)
Last-byte latency: 0.036 - 0.085 0.065 ( 0.015) 0.065 ( small objs) tx732aae54c9484689b8fea-0051b21709
INFO:Scenario run results saved to /tmp/ssbench-results/Small_test_scenario.u4.o613.r-.2013-06-07.102314.stat.gz
INFO:You may generate a report with:
.../ssbench-master report-scenario -s /tmp/ssbench-results/Small_test_scenario.u4.o613.r-.2013-06-07.102314.stat.gz
The default, textual table report may be seen in the above example output. You
can also specify --csv when running a scenario or generating a report later
to generate a CSV report instead. This feature is still pretty new so expect
the CSV report output to change over time.
Right now, the default report's CSV version is two lines: a line of column
header names and one line of actual data. Both lines are very long and the
set of columns present in any given CSV report will depend on the scenario
which was run. Some column names have the --pctile value in them and many
columns have the object sizes in them, which are defined in the scenario file.
You can think of the two CVS lines as a linear denormalization of the contents
of the two-dimensional table output.
Assuming the Swift cluster being benchmarked is not the bottleneck, the scalability of ssbench may be increased by
- Running up to one
ssbench-workerprocess per CPU core on any number of benchmarking servers. - Increasing the default
--batch-sizeparameter (defaults to 1) on both thessbench-masterandssbench-workercommand-lines. Note that if you are running everything on one server and using the--workersargument tossbench-master, the--batch-sizeparameter passed tossbench-masterwill be passed on to the automatically-startedssbench-workerprocesses. - For optimal scalability, the user-count (concurrency) should be greater than
and also an even multiple of both the batch-size and number of
ssbench-workerprocesses.
As a simple example, on my quad-core MacBook Pro, I get around 9,800 requests
per second with --noop (see below) with this command-line (a
--batch-size of 1):
$ ssbench-master run-scenario ... -u 24 -o 30000 --workers 3 --noop
But with a --batch-size of 8, I can get around 19,500 requests per second:
$ ssbench-master run-scenario ... -u 24 -o 30000 --workers 3 --noop --batch-size 8
When running ssbench-worker on a Mac, using HTTPS, I got a significant
speed-up when setting OPENSSL_X509_TEA_DISABLE=1 in the environment of my
ssbench-worker processes. I found this tip via a curl blog post after
noticing a process named trustevaluationagent chewing up a lot of CPU
during a benchmark run against a cluster using HTTPS.
To test the maximum throughput of the ssbench-master <==>
ssbench-worker infrastructure, you can add --noop to a
ssbench-master run-scenario command and the scenario will be "run" but
the ssbench-worker processes will not actually talk to the Swift cluster.
In this manner, you may determine your maximum requests per second if talking to the Swift cluster were free.
The reported "Average requests per second:" value in the "TOTAL" section of the report should be higher than you expect to get out of the Swift cluster itself.
With an older version of ssbench which used a beanstalkd server to manage
master/worker communication, my 2012 15" Retina Macbook Pro could get ~2,700
requests per second with --noop using a local beanstalkd, one
ssbench-worker, and a user count (concurrency) of 4.
With ZeorMQ sockets (no beanstalkd involved), the same laptop can get between
7,000 and 8,000 requests per second with --noop.
First, please use the Github Issues for the project when submitting bug reports or feature requests.
Code submissions should be submitted as pull requests and all code should be
PEP8 (v. 1.4.2) compliant. Current unit test line coverage is not 100%, but
code contributions should not lower the code coverage (so please include
new tests or update existing ones as part of your change). Running tests will
probably require Python 2.7 and a few additional modules like flexmock and
nose.
Regarding test tools, I started out using flexmock, but plan to mostly add
new tests using the mock library since that's been included in the stdlib
and the Python community seems to be converging on it. So please use mock
instead of flexmock for new tests.
If contributing code which implements a feature or fixes a bug, please ensure a Github Issue exists prior to submitting the pull request and reference the Issue number in your commit message.
When submitting your first pull request, please also update AUTHORS to include yourself, maintaining alphabetical ordering by last name.
If any of the file(s) you change do not yet have a copyright line with your name, please add one at the bottom of the others, above the license text (but never remove any existing copyright lines). Your copyright line should look something like:
# Copyright (c) 2013 FirstName LastName