src/

Build

1. Clone the copilot repository

git clone git@github.com:princeton-sns/copilot.git

1. Build

cd copilot
make

Run

1. Start the master service that waits for 3 replicas to connect

(./bin/master -N 3 -twoLeaders=true) &

1. Start 3 replicas that run Copilot protocol.

(./bin/server -maddr=<master-ip> -mport=7087 -addr=<node-0-ip> -port=7070 -copilot=true -exec=true -dreply=true) &
(./bin/server -maddr=<master-ip> -mport=7087 -addr=<node-1-ip> -port=7070 -copilot=true -exec=true -dreply=true) &
(./bin/server -maddr=<master-ip> -mport=7087 -addr=<node-2-ip> -port=7070 -copilot=true -exec=true -dreply=true) &

These replicas will connect to the master and wait for client connections. (Note: we assume the replicas are on separate machines. If they are on the same machine, each replica should use a different port number -port. To run Latent Copilot, replace -copilot=true with -latentcopilot=true.)

1. Start a Copilot closed-loop client with an id = 0

(./bin/clientmain -maddr=<master-ip> -mport=7087 -twoLeaders=true -id=0) &

For more options to specify for a server/client, please refer to startexpt.sh

Running Copilot on Emulab

System Requirements

We evaluated Copilot on the Emulab testbed using the d430 machines, each of which has one 2.4GHz 64-bit 8-Core processor, 64GB RAM, and is networked with 1 Gbps Ethernet. Each machine ran Ubuntu 16.04 STD. We built Copilot using a Go version go1.15.2 linux/amd64. We used a Python version 2.7.12.

Running startexpt.sh

On Emulab, after an experiment is created and swapped in, each node is assigned with an alias node-X (X >= 0). We provide an experiment script startexpt.sh that automates the process of starting the master, replicas, and clients. In this script, we assume the following setting:

• node-0: the control node where we invoke startexpt.sh.
• node-1: the master node.
• node-2,...,node-{n+1}: n replica nodes
• node-{2+n},...: client nodes
• Copilot source code directory and binaries are placed in a shared location accessible via a Network File System (NFS)

Usage: ./startexpt.sh <protocol> <num-replicas> <num-client-instances> <num-client-nodes>

For example, the following command will set up a Copilot cluster of 5 replicas and 30 closed-loop client instances sending their commands to these replicas. The client instances are evenly distributed across 10 nodes/machines.

# Run Copilot with 5 replicas and 30 clients (spread across 10 client nodes)
./startexpt.sh "copilot" 5 30 10

(Note: to run experiment with the Latent Copilot, specify latentcopilot instead of copilot. )

Experiment Data

When startexpt.sh is executed, a random ID is generated for this experiment, and a subdirectory that is named based on this ID is created under experiments/. This subdirectory contains raw data—e.g., throughput and latency—from each client, and the overall data that is output by the post-processing scripts.

• percentilesnew.txt: command latency (in microsecond) at every percentile
• tput.txt: system throughput (commands/second)
• tputlat.txt: system throughput and percentile latencies put together
• sys_tput.txt: instantaneous system throughput (interval is specified in tput_interval_in_sec in startexpt.sh)

Other Notes

Open-loop Clients

We also provide an implementation for open-loop clients, which can be found at src/clientol. An open-loop client sends commands in an open loop and the time between successive commands of a client follows an exponential distribution. The rate parameter (commands/second) of the distribution can be specified using the option -target_rps.

# Run an open-loop client with the arrival rate of commands is 2000 commands/second
./bin/clientol -maddr=<master-ip> -mport=7087 -twoLeaders=true -id=0 -target_rps=2000