Experiment Driver

Platform Setup

One node (or personal machine) to run the experiment driver. One cluster consisting of one loader node connected to one or multiple worker nodes via passwordless ssh (as in Cloudlab). The node running the experiment driver needs to have an ssh agent running to connect to the loader node. The trace files should be present on the node running the experiment driver. In the first phase, the driver will connect to the loader node and transfer the relevant trace and configuration files. It will then start the experiment, during which the loader node will send invocations to the worker node(s) and collect the results. The driver will then transfer those result files to its own node. After this it will send a signal to the loader node to clean up the state of the cluster, to prepare it for the next experiment.

Usage

Configuring the Driver

The following fields in driverConfig.json can be configured:

"username": Your username on the target Cloudlab cluster, used to ssh into the loader node. Note that you will need an ssh agent configured on the machine running the experiment driver.
"experimentName": The name of the experiment, creates a directory under that name within the specified output directory.

"localTracePath": The path to the trace files on the machine running the experiment driver. The trace files need to be called {n}_inv.csv, {n}_run.csv and {n}_mem.csv, with n being the number of functions. See further below for a more detailed explanation.

"loaderTracePath": The path at which trace files on the loader node should be stored, can remain unchanged.
"loaderAddresses": The hostname(s) of the loader node(s), for example "hp004.utah.cloudlab.us". Currently only one loader is supported.

The driver runs multiple experiments sequentially with different traces. Currently, this is based on the number of functions within the trace. For example, if you want to sweep traces with 50, 100, 150 and 200 functions, you would put the trace files within one folder, with the invocation files named 50_inv.csv, 100_inv.csv, 150_inv.csv and 200_inv.csv. The durations and memory trace files should be named {n}_dur.csv and {n}_mem.csv, with n being the number of functions. Then you would set the following parameters to sweep these trace files:
"beginningFuncNum": 50
"stepSizeFunc": 50
"maxFuncNum": 200
This will run a total of four experiments.

"experimentDuration": How long the measurement phase of the experiment should last. The total duration of the experiment is then the duration of the measurement phase + the duration of the warmup phase + 1 minute for the profiling phase.
"warmupDuration": The duration of the warmup phase.
"workerNodeNum": The number of worker nodes in the cluster
"outputDir": The local output directory on the machine running the experiment driver, this is where experiment results will be saved.
"YAMLSelector": Which yaml specification to use, supported values are "container" and "firecracker".
"IATDistribution": The IAT distribution that the loader will use when sending invocations to the worker node(s). Supported values are "exponential", "uniform" and "equidistant".
"loaderOutputPath": The output path for the experiment results on the loader node.
"partiallyPanic": Whether to enable partially panic mode. This modifies the panic-window-percentage from the default value of 10.0 to the maximum value of 100.0, and it modifies the panic-threshold-percentage from the default value of 200.0 to the maximum value of 1000.0.
"EnableZipkinTracing": Whether to enable Zipkin Tracing. Calls InitBasicTracer to initialise an OpenTelemetry tracer.
"EnableMetricsScrapping": Whether to enable Metrics Scraping.
"MetricScrapingPeriodSeconds": The interval at which metrics scraping occurs, in seconds.
"separateIATGeneration": Whether to generate IATs before running the loader. Leave as false unless you are running multiple loader nodes for one experiment (which is currently not supported). See issue #88.
"AutoscalingMetric": Which autoscaling metric Knative should use. Default is "concurrency", supported values are "concurrency" and "rps". See Knative Doc for more details.

Running the Experiment

$ go run experiment_driver.go -c driverConfig.json