progbg 0.2

A programmable way help quickly run and produce benchmarks and create graphs

progbg

The Programmable Benchmarker and Graphing Tool

The problem: I have to benchmark my system. I want to evaluate its success by using specialized benchmarks that run using it or on top of it. These workloads generate data which needs to be parsed into a usable, analyzable form. I then would like to generate figures to firstly to better see trends in my data as well as present findings to others.

Solution: progbg is a simple glue framework that exposes a planning API to allow you to plan and program your benchmarks, parsing, and finally the graphs and figures that are produced by them. The API is developed to allow not just basic option paramaters to modify graphs and benchmarks, but also defining and changing behaviour through defined functions and classes that have access to matplotlib which progbg uses for its graphing library.

Installation

progbg can be installed through pip

$ pip install progbg

Quick Start

Firsty lets define the following terms:

1. Execution are composed of backends and a benchmark. They form the running component of progbg
2. Backend: The system underlying system you may be trying to test. This is setup prior to running a benchmark. Benchmarks can be composed using path style syntax (See below for an example).
3. Benchmark: An application that runs on a backend and produces data output, using a given parser.
4. Parser: Parsers that consume data to produce objects which graphs use.
5. Graph: Consume parser information and formatting information, to help output individual graphs. Can be used by themselves or be used by figures.
6. Figure take formatting information (styles etc.) and are composed of graphs.

Simply put you define benchmarks and backends with the provided decorators registerbackend and registerbenchmark these wrap classes.

Sample Backend would look like:

@sb.registerbackend
class FFS:
    @staticmethod
    # You can use whatever names for arguments you'd like but arguments should be
    # a unique name
    def init(myvar = 10, pass_me_in = 5):
        print("FFS Backend initing {} {}".format(myvar, pass_me_in))

    @staticmethod
    def uninit():
        pass

# Another backend as well
@sb.registerbackend
class BCK:
    @staticmethod
    def init(another_var = 5):
        print("ANOTER Backend initing {} ".format(another_var))

    @staticmethod
    def uninit():
        pass

Backends require both the init and uninit methods, these are what are ran before benchmarks are executed.

A sample benchmark would look like the following:

@sb.registerbenchmark
class WRK:
    @staticmethod
    def run(backend, outfile, test = 5, x = 10):
        # DO STUFF
        avg = str(randint(1, 1000))
        max = str(randint(500, 1000))
        min = str(randint(1, 499))
        vals="{} {}\n".format(test, x)
        strn="Latency  {}  {}  {}".format(avg, max, min)
        with open(outfile, 'w') as out:
            out.write(vals)
            out.write(strn)

This is just a dummy benchmark that randomly produces 3 numbers.

Note:

1. Benchmarks require at least 1 argument which is that path to an output file, which the benchmark should use to pipe output related to the data from the benchmark. This can be seen as the path for the -o argument many benchmarks use. If a benchmark does not use that. I suggest using subprocess and piping the output to the file.
2. All variable names used for arguments within methods for backends or benchmarks require a name, and to be unique within the scope of the entire file. For example above, the WRK benchmark and FFS backend cannot both have a variable named test.

Now that the backends and benchmark is defined, we can use these to define our execution.

A simple execution:

sb.plan_execution("my_execution",
        # Notice the "wrk" - this is to identify which registered benchmark to use (specified above)
        sb.DefBenchmark("wrk",
            sb.Variables(
                consts = {
                    "test" : 2
                },
                var = [("x" ,range(0, 3, 1))]
            ),
            iterations = 5,
            parse = sb.MatchParser("out",
            {
                "^Latency": (
                    ["avg", "max", "min"], func
                )
            },
        )
        ),
        {
            # Note: The arguments passed in here have the same name as the arguments used
            # for registered benchmark
            'ffs/bck': sb.Variables(
                consts = {
                    "myvar": 5
                },
                var = [("pass_me_in", range(0, 10, 2))]
            ),
            'bck': sb.Variables(
                consts = {},
                var = [("another_var", [1, 5 ,7])]
            )
        },
        out = "out.db"
)

Planning Executions

plan_execution takes in 3 arguments:

A unique name

This is a name that will be used by graph objects te help create graphs. It is used to identify the execution and all data produced by it.

A Benchmark object

Currently there is only one - DefBenchmark. This benchmark required 4 parameters, the lowercase name of the registered class, its Variable objects, number of iterations that you'd this benchmark to run, and the parser that will convert data from the benchmark to usable objects.

Output string Currently there are two supported outputs for benchmarks. If the user supplies a name with a ".db" extension, then data will be placed into a sqlite3 database. Multiple executions can share this database.

Any other name will be seen as a directory and each benchmark run will be stored as a seperate file.

Variables Objects

These objects define what changes and stays constant during the execution run. They are used within both defining variables for the backend as well as the benchmark. This allows progbg to iterate through the cross product of all changing variables defined within the backend and the benchmark. For our benchmark we are changing the variable "x" and keeping "test" constant. The given backends are also changing their own variables. These variable names are the arguments to the benchmark and backend registered above.

A Dictionary that describes the backends

These are the backends you wish to run. Notice the format to describe them. When we register backends we can refer back to them using their lower case name, and we are also able to compose these backends.

"ffs/bck" uses a composition of both the FFS defined backend and BCK backend (found above). This inits the backend in the order of the path. With the benchmark being run after all defined backends init function have been called. This for example allows you to create a fast file system backend, then place some server (tomcat, or whatever you need) on top then run the composed benchmark.

Note that backends are created and destroyed after EVERY benchmark, this can be disabled with the --no-reinit flag.

Match Parser

The Match parser takes a dictionary of regex strings to a tuple of named outputs and a function that creates those outputs. When the parser receives raw data it goes through line by line, and passes the line to the provided function if the regex search is true on that line

For our above example this is matching on the regex "^Latency" and outputing our variables avg, max, and min. These can be used by the graph tool to be graphed. The names must also be unique globally.

Planning Graphs

Finally we would like to produce a graph (or many) from our executions.

sb.plan_graph("graph-1",
    sb.LineGraph(
        "x",
        "avg",
        ["my_execution:ffs/bck"],
        {
            "pass_me_in": 4,
            "another_var": 5,
            "test" : 2,
        },
        out = "avg.svg"
    ),
)

We use the plan_graph function, which takes 2 arguments:

Unique Name

This name is used so that we can plan figures if we so desire with these graphs

Graph Object

Currently there are two graph styles (LineGraph and BarGraph). Refer to docs for more detailed usage. But in general, these graphs take the variables that you wish to graph by their name (these are names that were provided either to the variables for the backend, the benchmarks, or the parser), and uses this to grab the appropriate data from the specified saved output (either the directory. or sql database).

In our above example we are plotting how "avg" changes over "x", and we want the variables pass_me_in, aother_var, test to be fixed at specific points. We then wish this graph to be outputed to avg.svg. The graph will always output an svg by default (for use within the flask server that can be used to present graphs).

progbg also supports .pgf extensions

More

Take a look at a working example plan. Try running it with the command progbg plan.py -p 8080 and go to localhost:8080 to view the example graphs!

Viewing Graphs

Graphs are saved by default in the graphs directory of the current working directory.

By passing the argument -p PORT to progbe, this will expose figures and graphs to be viewed by your browser at http://localhost:PORT

Author

Kenneth R Hancock krhancoc <https://github.com/krhancoc>

TODO

1. CSV Parser
2. Command line Benchmark Class (We may also not even need this, we may only need one benchmark class)
3. Better Error Handling and Hints
4. Tests