CsharpRAPL

CSharpRAPL is a framework for benchmarking C# in regards to energy.

In the Benchmarks folder, all the benchmarks created for this project can be found.

CSharpRAPL contains the library code we have implemented to use for benchmarking.

This code is a continuation and extension of lrecht/ParadigmComparison, which is an earlier university project.

By default, the CSharpRAPL tries to make each loop iteration take 2 seconds called Dynamic Loop Iteration Scaling.

CSharpRAPL also has Dynamic Iteration Calculation which scales the number of loop iterations according to the deviation of the results.

The results generated is a CSV file that contains the DRAM Energy, Temperature, Elapsed Time, Package Energy, and the return value of the Benchmark.

Note that the separator is ; and not ,

Example

ElapsedTime;PackageEnergy;DramEnergy;Temperature;BenchmarkReturnValue 0.31062468886375416;3341.3842320442195;172.80355095863342;43;19 0.310964733362198;3335.9728753566737;172.84899950027466;43;19 0.31095251441001887;3353.798389434814;173.25878143310547;43;19 0.31064271926879894;3366.4405345916743;172.3937690258026;43.5;19

The results can be visualized using Box Plots an example of this can be seen on the following image (Note that the following image is only for illustrative purposes and the results might not be correct):

Requirements

• Linux with access to RAPL
• libgdiplus - Is required for plotting
• cpuset - Required for creating a cpuset to execute benchmarks. Can be used without.

Usage

Each of your benchmark class must contain the following fields:

public static ulong Iterations; public static ulong LoopIterations;

This is needed to make use of Dynamic Iteration Calculation and Dynamic Loop Iteration Scaling.

Note that both ulong and unit is supported.

Registering Benchmarks

There are two ways of registering benchmarks, either using Attributes or doing it manually. The most tested method is using Attributes.

Attributes

We can register a benchmark by adding the Benchmark attribute where the first argument is the benchmark group and the second is the description for the benchmark. An example can be seen below.

[Benchmark("Addition", "Tests simple addition")] public static int Add() { int a = 10; int res = 0; for (int i = 0; i < LoopIterations; i++) { res = a + i; } return res; }

The attribute also has options for order and if it should skip the benchmark.

To take advantage of using the Attributes you have to use the BenchmarkCollector which collects all methods with the Benchmark attribute.

When using the attributes you can also make use of Variations attribute, which creates variations of a field or property.

A VariationBenchmark must be present on the methods which uses the variations. This is to avoid creating lots of extra variations that aren't needed.

Example:

[Variations(10, 15)] public int TestProp { get; set; } [Variations(21, 12)] public int TestProp2 { get; set; }

Here we have two properties which will create 4 variations like seen below:

Variation 1: TestProp = 10, TestProp2 = 21 Variation 2: TestProp = 10, TestProp2 = 12 Variation 3: TestProp = 15, TestProp2 = 21 Variation 4: TestProp = 15, TestProp2 = 12

So it will create an instance of the benchmark where each these values are set according to the table above.

Another attribute that is available is the TypeVariations attribute which is used for creating benchmark variations where we want to use different instances of a class like so:

[TypeVariations(typeof(ArgumentException), typeof(ArgumentNullException), typeof(ArgumentOutOfRangeException))] private Exception Exception = new Exception();

The above code if used will create 3 variations where in each variation Exception is assigned to a new instance for each of the types in the attribute.

The field/property should have a default value as to not give null-reference expectations when running the no-variation benchmark.

To use this the method must be marked with VariationBenchmark attribute.

Manual

You can add benchmarks manually by either using BenchmarkCollector or BenchmarkSuite to do that use the RegisterBenchmark method.

Example of manual registration:

public static void Main(string[] args){ var benchmarkSuite = new BenchmarkSuite(); benchmarkSuite.RegisterBenchmark("TestGroup", DummyBenchmark); } public static int DummyBenchmark() { return 1; }

Running the Benchmarks

To run the benchmarks simply call RunAll on either your BenchmarkSuite or BenchmarkCollector instance.

For more examples look at Suite.cs or at the tests.

Scripts

We have provided the scripts:

• Benchmarks/execute.sh
  • Sets the CPU scaling governor to performance
  • Executes the benchmarks on shielded cores (see cpuSet.sh)
  • Sets the CPU scaling governor to powersave after running the benchmarks
• Benchmarks/removeResults.sh
  • Removes old results both in the form of CSV and plots. (Alternatively use -r)
• Benchmarks/zipResults.sh
  • Zips CSV results into a zip file. (Alternatively use -z)
• Scripts/cpuSet.sh
  • Shields cores number 2 and 4 (0 indexed) from other processes, including the kernel. Uses the ubuntu package "cpuset", which is a Python wrapper over the Linux cpuset interface. This shield is to be used when executing benchmarks.
• Scripts/runAtStartup.sh
  • First makes sure hyperthread cores are disabled, then sets swappiness to 10, disables ASLR, lowers perf sampling interval, and disables Intel turbo-boost for consistency.
• Scripts/setPerformance.sh
  • Sets the cpu scaling governor to performance.
• Scripts/setPowerSave.sh
  • Sets the cpu scaling governor to powersave.

CLI Options

-g, --SkipPlotGroups If plotting each benchmark group should be skipped. -i, --Iterations (Default: 0) Sets the target iterations. (Disables Dynamic Iteration Calculation) -l, --LoopIterations (Default: 0) Sets the target loop iterations. (Disables Dynamic Loop Iteration Scaling) -k, --KeepOldResults If not set removes all files from the output folder and the plot folder. -o, --OutputPath (Default: results/) Set the output path for results. -p, --PlotResults Should the results be plotted? -a, --BenchmarksToAnalyse The names of the benchmarks to analyse. -z, --ZipResults Zips the CSV results and plots into a single zip file. -j, --Json Uses json for output instead of CVS, includes more information. -m, --CollectMemoryInformation Collects memory information before and after each benchmark. --TryTurnOffGC Tries to turn off GC during running of benchmarks. --GCMemory (Default: 250000000) Sets the amount of memory in bytes allowed to be used when turning off garbage collection. --PlotOutputPath (Default: _plots/) Sets the output path for plots. --OnlyPlot Plots the results in the output path. --OnlyAnalysis Analyse the results in the output path. --Verbose Enables debug information. --OnlyTime Only measures time. --help Display this help screen. --version Display version information.