Running.md

Running the Benchmarks

Parameter Space

The experiments have two options for parameter spaces:

1. The "full" parameter space is targeted at machines with 30-50 CPU cores, and is the default setting.
2. The "testing" parameter space is targeted as machines with 2-8 CPU cores, and can be selected by passing --testing true to ./bench.sc in one of the execution modes listed below.

Config

• Check the javaOpts in benchmarks.sc and make sure the the -Xmx matches available memory on your system, so that the experiments do not start paging. Divide the memory between the number of nodes you are planning to run in fakeRemote!
• If running remote, set masterAddr in bench.sc to an IP address (and port) that all nodes in nodes.conf can reach.

Local Execution

Run ./bench.sc local to only run the benchmarks that execute within a single process.

Local Distributed Execution

Run ./bench.sc fakeRemote to run all benchmarks on the local machine, by starting a configurable number (--withClients) of clients in temporary folders.

Remote Distributed Execution

Write all client nodes into nodes.conf. Make sure the master can SSH into them without a password, and they can communicate over TCP on all ports. Then run ./bench.sc remote to run all benchmarks over the configured set of nodes.

Other Options

• A subset of implementations X,Y,Z can be selected by adding --impls X,Y,Z to the ./bench.sc command.
• A subset of benchmarks X,Y,Z can be selected by adding --benchmarks X,Y,Z to the ./bench.sc command.

Plotting

Results can be automatically plotted by running ./plot.sc. It requires GNUPlot to be installed.