Run performance tests on network software by running snippets of bash scripts on a cluster following a simple definition file. For instance, the following configuration to test iPerf3 performance (omitting graph styling options):

%info
IPerf3 Throughput Experiment

%variables
PARALLEL=[1-8]
ZEROCOPY={:without,-Z:with}

%script@server
iperf3 -s &> /dev/null

%script@client delay=1
result=$(iperf3 -f k -t 2 -P $PARALLEL $ZEROCOPY -c ${server:0:ip} | tail -n 3 | grep -ioE "[0-9.]+ [kmg]bits")
echo "RESULT-THROUGHPUT $result"

Will automatically produce the following graph:

When launching npf:

npf-run --test tests/tcp/01-iperf.npf

Test files allow to define a matrix of parameters to try many combinations of variables for each test and report performance results and evolution for each line of the matrix.

Finally, a graph will be built and statistical results may be computed for each test showing the difference between variables, different softwares, or the evolution of performances through commits.

Test files are simple to write, and easy to share, as such we encourage users to share their ".npf" scripts with their code to allow other users to reproduce their results, and graphs.

NPF supports running the given test across a custer, allowing to try your tests in multiple different configuration very quickly and on serious hardware.

NPF is built using Python 3, and is published on pypi, so it can be installed with pip using:

pip3 install --user npf

Your .npf test file is composed of a serie of sections, as in the example given above. The sections describe the scripts to run, where to run them, what variables should be tested, what are their ranges, configuration parameters such as timeout or graph colors, etc. Each section is described in more details in tests/README.md. When launching NPF, you will also give the name of one or more repositories, which are files located in the repo folder describing software to download, install and compile so everything is in place when your experiment is launched. They follow a format descrived in repo/README.md. Your test script will also define a few script roles, such as client or server as in the example above. When you actually launch your experiment, you must tell which machine (physical or virtual) will take the role. For simple cases, passing the address of a machine with the --cluster role=machine will be enough. When you'd like to define parameters such as IPs and MAC addresses, you can define a cluster file that will describe details about each machines. See cluster/README.md for more details.

Cluster-based tests use SSH to launch multiple software on different nodes, therefore SSH should be setup on each node for a password-less connection. Use public key authentication and be sure to add the ssh keys in your ssh agent using ssh-add before running NPF.

Most DPDK-based but also other scripts use the sudo=true parameter in test scripts to gain root access. You can either always connect as root to other servers (see the cluster section below) or set up password-less sudo on all nodes.

Use either a NFS shared mounted on all nodes or sshfs to mount the local NPF folder on all nodes. The path to the shared NPF root can be different on each node, see the cluster section below. If this is not the case, the dependencies (software built by NPF) will be sent to all nodes that will use them in the corresponding scripts through SSH, but it is slower.

Three tools come with this performance framework :

• npf-run for advance regression and statistics tests on one repository
• npf-watch to watch one or multiple repositories for any new commit and e-mail regression results in case of change in performances due to the last commits
• npf-compare to compare one test script but across multiple repository, mainly to compare how different branches/implementations behaves against each others

NPF-Run is the main NPF tool.

It checkouts or updates a given repository (described in the repo folder), build the software, and launch the given test scripts

Example :

    npf-run click --test tests/click/ #Produce a graph for each click-based tests with the result

NPF-Run is able to check commit history, do regression test, and graph the performance history for all test script using the --regress flag.

    #click master is updated
    npf-run click --test tests/click/ --regress #The graph now compares HEAD and the last commit, if major performances changes are found, the return code will be different than 0
    #click master is updated again
    npf-run click --test tests/click/ --regress #The graph includes the older commit for reference, up to "--graph-num", default is 8

Example of a generated graph for the Click Modular Router, just when IPFilter compilation process was re-worked :

Alternatively, you can force npf-run to re-build and compute the data for the old runs directly with the --allow-old-build option :

    npf-run click --test tests/click/ --allow-old-build [--graph-num=8] #Graph the performance of the current version and the last 8 previous ones

NPF-Run can produce statistics about the results such as the best set of variable, the average per-variable, a regression tree and the importance of each features.

    npf-run click --test tests/click/ --statistics

See npf-run --help for more options

Watcher is a stripped down version of npf-run (without statistics support mostly), but allowing to loop through a given list of repositories watching for changes. When a new commit is seen, it will run all given test scripts and e-mail the results to a given list of addresses.

npf-watch click fastclick --mail-to barbette@kth.se --tags fastregression --history 1

The arguments are :

• click fastclick : List of repos to watch, as described in the repos folder
• --history N allows to re-do the tests for the last N commits, you will receive an e-mail for each commits of each repos.
• --tags TAG1 TAG2 [...] allows to set flags which change variables in the tests, see below.

See npf-watch --help for more options

NPF-Compare allows to do the contrary of npf-run : instead of testing multiple npf scripts on one repository, it tests one test script across multiple repositories.

This example allows to compare Click against FastClick (a faster version of the Click Modular Router) for the infinitesource test case :

npf-compare click fastclick --test tests/pktgen/infinitesource-01.npf --variables LENGTH=64

• click fastclick : List of repos to compare
• --test FILENAME : Test script to test. This argument is available in all tools.
• --variables VAR=VAL [...] : Fix the value of one variable. By default in this test script, the test is redone with packet length 64,256 and 1024. This allows to have one less "dynamic" variables so the grapher can use a lineplot instead of a barplot (see below).

Result : Just for relevance, batching is what makes this difference.

This tool has also less options than NPF-Run. You should use NPF-Run to create your tests and tune parameters for each repository independently. And then only use npf-compare when ready. Compare does not support statistics or regression tests.

See npf-compare --help for more options

Use npf-run for development, trying big matrices of configuration, get extended graph and customized tests for each test scripts.

Use npf-watch with the fastregression tags to send you an e-mail automatically when some new commits introduce performances problems.

Use npf-compare to compare multiple repositories, multiples branches or multiple different softwares. The test scripts included in this repository support comparing throughput of Click and FastClick in diverse configurations, or NetPerf and Iperf as packet generators.

All tools feature those common arguments :

• --variables VAR=VAL [VAR2=VAL2 ...] allows to overwrite some test script variables configuration, mainly to reduce the set of parameters
• --config VAR=VAL [VAR2=VAL2 ...] allows to overwrite some configuration options such as n_runs used to define the number of time a test should be launched for each variable combination
• --test path : Path to a folder or a test script. By default "tests" is used.

All results of tests are saved per-commit and per-repo to avoid re-computing the next time you launch either of the tools. However the set of variables must be exactly the same.

To force re-computing the tests, use the --force-test option. The --force-rebuild may also be something you want.

You can customize the output of NFP by passing different arguments. Below, you can find some of the common outputs:

• --pandas [PATH] NFP produces a single Pandas dataframe if you use this argument. Later, you can load the dataframe for post-processing. The following code shows one example for a sample dataframe (i.e., test-pandas.csv) with two variables (i.e., X and Y). Line 3 produces the median of multiple runs, while line 4 shows values of all runs in a list, which can be used for a boxplot.

1 import pandas as pd
2 df = pd.read_csv("test-pandas.csv")
3 df[['X','Y']].groupby('X').agg({'Y' : ['median']})
4 df[['X','Y']].groupby('X').agg({'Y' : lambda x : list(x)})

• --output Outputs a standard CSV for each output variable. According to --output-columns, by default the X variable(s) and the average of Y value. For instance --output-columns x perc1 perc25 median perc75 perc99 avg would have 7 columns with the X value, then the 1st percentile of the results for the variable, etc. For instance, if you have a variable PARALLEL=[1-8] and you collect THROUGHPUT, for which you did 3 runs, by default you'll get a csv file for THROUGHPUT, that gives you :

  1 2.2
  2 3.4
  3 4.2
  ...
  

  Where the second column is the average of the 3 runs. Instead, with the columns "x all", you would get:

  1 2.0 2.2 2.4
  2 3.1 3.4 3.7
  3 4.0 4.2 4.4
  ...
  

All programs have the --tags argument, allowing to give a set of tags that trigger changes in the test scripts. The dpdk flags tells that a DPDK environment is setted up with at least two NICs, allowing DPDK-based tests to run. The fastregression tag allows to only try important variable combination and not search for performance points, while full is the contrary and will run a very big set of variables combinations to get statistics out of results.

See tests/README.md to read more about test scripts and learn about creating new ones. Scripts describe the tests and parameters to re-run them in multiple configuration. Test scripts are the heart of NPF.

See repo/README.md to lear how to build repository definition files to let NPF know how to fetch and compile some software

Test scripts define multiple roles such as "client" or "server". Each role can be mapped to a given node to run a test across a cluster using the --cluster ROLE=NODE [ROLE2=NODE2] argument.

NPF will run the test scripts for each role on the mapped cluster. Giving the node address in the command line may be enough. However some tests require more information about each node that can be set using cluster files. More information about writing cluster files is given in cluster/README.md

Graph are automatically generated for all tested variables combinations.

To choose the type of graph, the number of dynamic variables is taken into account.

Below, npf-run gave two series to the Grapher (current and last commit), while the test script generate a matrix of Burst and Lengths, that is 2 dynamic variables and only a barplot can render that correctly as lines would be uncomparable.

If a "previous version" is not given to npf-run (so it just runs the test for the current master but do not compare the results), the graph will use one variable as the serie as having only one line would be a loss of space, leaving only one dynamic variable :

The Comparator uses the repositories as series.

The graph can be tweaked using many options, and the data can also be transformed using multiple tools to better display results. Data transformation will also affect the output CSV. In any case none of these options affect the values in the data cache, so you may try different tweaks without risks.

All the following options can be added to the %config section of the test script, or after the --config parameter on the command line of any of the tools.

This section is in rework.

• var prefix generally affect variables. It generally takes a list of variables, or a dict of variables->parameters.
• graph prefix only affect the graph, styling ,etc. The CSV data will not be changed.

• graph_error_fill=true/false Display a "filling zone" instead of error bars. To be used when you have many points and your graph becomes horrible.

• graph_color={0,1,2,3} Select a set of color for each serie. Colors are predefined ranges. Default is to use the 0 set of colors for all series, the serie 0 is a mix of different colors, while 1 to 5 are shades of the same colors.
• graph_markers={'o', '^', 's', 'D', '*', 'x', '.', '_', 'H', '>', '<', 'v', 'd'}, type of markers to be used, per-serie. Default if this. See matplotlib docs to find the type.
• graph_lines={'-', '--', '-.', ':'} Type of lines, per-series.

• var_log, [THROUGHPUT,LATENCY] Define variables that should be shown in a log axis in base 2
• var_log_base, {THROUGHPUT:2} Define variables that should be shown in a log axis, with the specified base (default is 2 with var_log)
• var_lim, {THROUGHPUT:0-100} Define the range for some variables, useful to cap graphs. One can produce "broken axis" graphes by giving multiple ranges separated by a + sign. For instance {THROUGHPUT:0-10+50-100}.
• var_format={THROUGHPUT:%dGbps} Printf like formating of va And the prefix should be changed.riables. Eg this example will display all visualisation of the value of throughput (eg in the axis) as XGbps. Use in combination to var_divider.
• var_ticks, {THROUGHPUT:0+5+10+15+20} Define where the ticks should be set, in this example there will be ticks in 0,5,...20.

• var_names, {"result-LATENCY":"Latency (µs)","result-THROUGHPUT":"Throughput"})
• var_unit, {"result": "bps","result-LATENCY":"us","latency":"us","throughput":"bps"})

• graph_scatter=true/false Use a scatter plot instead of a lineplot, default false. You must arrange the data so it displays as a line plot (one dynamic variable only).
• graph_grid=true/false Display a grid on the graph. Default false.
• graph_bar_stacks=true/false If your series are a complex barplot (more than 1 dynamic variable), it will stack the plots instead of adding them one after the other. Default is false.

• graph_series_sort=method Sort series according to the method wich can be : "natsort", natural alphabetical sorting, "avg", "min" or "max" to sort according "y" values. The sorting can be inversed by prefix the method with "-". Default is to not reorder.
• graph_max_series=N limint the number of series to N, used in conjunction with graph_series_sort to only show the "best" series. By default there is no limit.
• graph_serie=variable Use a specified variable as the serie of a line plot.

• graph_legend=true/false Enable/disable legend. Default is true.
• title=title Title of the graph
• var_hide={A,B,...} List of variables to hide
• var_label_dir={A:vertical,B:horizontal} Force the direction of labels on the X axis for the given variables. By default, vertical when there are more than 8 values, horizontal otherwise. Accepted values: vertical, horizontal, diagonal
• graph_force_diagonal_labels=true/false Always use diagonal labels for the X axis, independently from the var_label_dir setting. False by default.

• var_combine={NUMA+CORE:SCORE} will combine multiple variables in a single one. Eg if you have a NUMA={0,1} variable, and CORE=[1-4] this will combine them as a single variable SCORE={0-1,0-2,0-3,0-4,1-1,1-2,1-3,1-4}. This allows to reduce the number of variables to graph, eg you may prefer to have a lineplot of SCORE, instead of a barplot of NUMA and CORE according to the serie.
• series_as_variables=true/false Will convert the series as a variable. This is useful in npf-compare to consider the different tags/software you used as a variable, and use something else as a serie.
• var_as_series={QUEUE,POLICY} list of variables to use as series. If multiple, or a serie already exists, it will do the cross product of the variables. Usefull to pass "trailing" dynamic variables as more lines in a lineplot.
• result_as_variables={COUNT-Q(.*):QUEUE-COUNT} Group multiple results matching a regex as a single variable. Eg if you run a single test that outputs multiple statistics for "the same thing", like the number of bytes per NIC queues, you will have your scipt display RESULT-COUNT-Q0 A, RESULT-COUNT-Q1 B, ... and this example will make a variable QUEUE with all the observed values, and create a new result type called "COUNT".
• series_prop=true/false Divide all results by the results of the first serie. Hence graphs will be a percentage of relative to the first series. Eg if the first serie is "software 1" it will be removed from the graph and the other series will show how much better software 2, ... did against software 1.
• var_divider, {'result':1}) Divide the variables or results by the given value.
• graph_map,
• graph_cross_reference, {Y:VARIABLE}, change the graph where the Y axis is Y (the result name) to have the X variable being another variable
• var_aggregate, {VARIABLE:method}, aggregates all values for a given variable. If "method" is "all", all results will be put in a single variable value like if they were all points for the same run. You can also use "median", "average", ... to combine results for all variables using those mathematical methods.

• graph_subplot_results={THROUGHPUT+LATENCY:2} combine two results in a single figure. If graph_subplot_type is subplot, then it will simply combine the graphs in a single file using the given number of columns. If the subplot_type is axis, it will use a dual-axis plot. Only the last variable will be on the second axis, so one may combine multiple variables on the same axis, like TX and RX throughput on the left axis, and the latency on the right axis?
• graph_display_statics=true/false Will add a subplot to show the value of static variables. Useful to exchange graphs with colleages knowing what are the fixed parameters for all the graph you show. But the results is somehow horrible.
• graph_text=string Add some texts under all graphs.

Have you read tests/README.md? Then, inspire yourself from the test script files in tests/click mostly, then write your own!

We welcome merge requests for generic stuffs! But you can keep your files in your "experimentation" folder. Indeed, NPF will always look for a file first in "./repo" for repo files, "./modules" for modules and "./cluster" for machines definition.