Concord: An efficient runtime for microsecond-scale applications

This repository contains the source code for Concord, which was published at SOSP'23. In particular, it contains a) the two LLVM passes used by Concord to automatically instrument applications and b) the implementation of the Concord runtime on the Shinjuku dataplane OS.

Organization

Subdirectories have their own README files.

• benchmarks - The different applications we tested Concord on and overhead measurements.
• concord-shinjuku - The Concord runtime integrated into the Shinjuku dataplane OS.
• figures - Scripts to generate the figures in the paper
• scripts - General scripts to run concord
• schedsim - Our queueing simulator
• src - LLVM passes and the Concord runtime library

Getting started

Clone the repository, please make sure you have cloned via ssh, otherwise the submodules will not be cloned.

git clone --recurse-submodules [email protected]:dslab-epfl/concord.git
cd concord

Then setup Concord, this script will build concord-cache-line, concord-rdtsc pass, and install the required dependencies.

./setup.sh

Simple Examples

First run the basic examples to see cache-line and rdtsc based instrumentation in action.

cd benchmarks/basics
make

See the instrumentation outputs:

> Module Name: <stdin>
> Enable instrumentation: 1
> Modified subloops: 200
> Disable bounded loops: 0
Instrumenting loop in function: hello
Instrumenting loop in function: main
> This is declaration, skip atoi
> This is declaration, skip puts
> Unique loops: 0

After run the programs, first run cache-line based instrumentation:

./hello.out

# Program Modified !! 
# Hello, world!

Then run rdtsc based instrumentation, it will print the number of cycles between two concord_rdtsc_func calls.

./hello_rdtsc.out
# 16070
# 16054
# 16060
# 16056
# 16060
# 16056
# 16060
# 16056
# 16060
# ...

To run more complex examples, including the entire runtime on Shinjuku, please refer to the README within the concord-shinjuku directory.