ALPS: An Adaptive-Learning, Priority OS Scheduler for Serverless Functions

Overview

Welcome to ALPS scheduler project. Our innovative kernel scheduler is designed to enhance the performance of Function-as-a-Service (FaaS) workloads, which are known for their ephermeral, highly concurrent, and bursty nature. Existing OS schedulers, such as Linux Completely Fair Scheduler (CFS), often fail to meet the unique demands of serverless functions, particularly those with short execution time. ALPS addresses this challenge by approximating the principles of the Shortest Remaining Process Time (SRPT) with the robust framework on CFS, delivering a dynamic, application-aware OS scheduling solution.

Getting Started Instructions

For ATC'24 Artifact review, to save reviewers’ effort, we strongly recommend that reviewers use our dedicated development machine. We will provide support to ensure that reviewers have full access to the machine throughout the artifact review process.

Operating System required

We have implemented ALPS based on Linux kernel version 5.18-rc5. You must compile and run the ALPS on the the kernel. We recommend build the kernel based on Ubuntu 22.04 LTS.

Software Required

The exact software used to build ALPS as follows:

• gcc
  • version: 11.4
• go
  • version: 1.21.10
• bpftool
  • version: 5.14.0
• schedtool
  • version 1.3.0

In addition, We modify and provide exact softwares binaries to run FaaS service.

• Docker
  • Docker client
    • version: 20.10.25
  • dockerd
    • version: 20.10.25
  • runc
    • version: 1.1.10
  • containerd:
    • version 1.6.24
• OpenLambda:
  • commmit hash: 92fbdfe

Step-by-Step Installation

gcc installation

apt update -y
sudo apt install build-essential
gcc --version 

go installtion

Download the Go language binary archive

wget https://go.dev/dl/go1.21.10.linux-amd64.tar.gz
sudo tar -xvf go1.12.linux-amd64.tar.gz
sudo mv go /usr/local

Setup Go environment, including GOROOT and GOPATH. Add environment variables to the ~/.profile.

export GOROOT=/usr/local/go
mkdir $HOME/project
export GOPATH=$HOME/project
export PATH=$GOPATH/bin:$GOROOT/bin:$PATH

Verify installation

~$ go version
~$ go version go1.21.10 linux/amd64

bpftool

Clone bpftool repository and build following installation instruction

git clone --recurse-submodules https://github.com/libbpf/bpftool.git

schedtool

Install schedtool by apt

sudo apt-get update -y
sudo apt-get install -y schedtool

Docker

Clone ALPS repository and copy docker binaries to /usr/sbin

git clone https://github.com/fishercht1995/ALPS.git
cd docker_binaries
cp binary-client/* /usr/bin/
cp binary-daemon/* /usr/bin/

OpenLambda

Build OL worker image

cd experiments && make imgs/lambda

We modify openlambda to support ALPS JSON configuration schema. We provide the ol binary and an example of configuration file at experiments/. In addition to OL configuration, ALPS allows user to define function meta to distinguish scheduling in the kernel as follow:

"seal_priority": "Function UID",
"function_name" : "Function Name",

Hello-Word Example

Start Docker service

cd experiments && ./docker.sh

Run ALPS frontend scheduler

cd frontend && python main.py

Build and Run AlPS backend

mv LINUX_SOURCE /Linux
cd backend
make bpf && make alps
./alps.o

Run Openlambda

cd experiments && ./ol worker --path={PATH}

Now invoke your lambda

curl -X POST http://localhost:5002/run/fib \
     -H "Content-Type: application/json" \
     -d '{"n":"30", "id":"20", "job":"fib"}'

Detailed Instructions

To test performance of ALPS, we provides some helper scripts, firstly, start and delete multiple functions:

cd experiments
./init_function.sh # init functions
./delete_function.sh # delete function workers

Then start frontend and backend.

cd frontend && python3 main.py --alpha 1 --beta 1 --ml avg --exp_result ../experiments/seals
cd backend && ./alps.o

Submit benchmark request by http client

go build run.go
cd http_client && ./test.sh 

Ablation Study

The frontend scheduler offer two parameter to control policy fine-tuning. By defauly, policy fine-tuning is enable

python3 main.py

And users can disable unpredictability finetuning and overload finetuning by

python3 main.py --unpred --overload 

Sensitivity Analysis

Users can control policy parameters

python3 main.py --alpha 1 --beta 1 --theta 50 --gamma 1

Users can also change machine learning methods by

python3 main.py --ml avg
python3 main.py --ml LR

Test for different workloads

To test different trace workloads, replace the workload file (exp) in the http_client folder. By default, the workload is generated by Huawei trace, and we offer an Azure-generated workload (azure) in the same folder.