DeepScaler: Holistic Autoscaling for Microservices Based on Spatiotemporal GNN with Adaptive Graph Learning
- Overview
- I. Machine Prerequis
- II. Installation
- III. Deploy Benchmark Microservices
- IV. Workload Generation
- V. Train and Test
- VI. Autoscaling
- VII. Evaluation
- VIII. Citation
- IX. Contact
This repository contains a prototyped version of DeepScaler described in our ASE '23 paper "DeepScaler: Holistic Autoscaling for Microservices Based on Spatiotemporal GNN with Adaptive Graph Learning".
| Aspect | Details |
|---|---|
| Operating System | Ubuntu 18.04 LTS |
| Kernel Version | 4.15.0 |
| VM Specifications | 4 with 12-core 2.2 GHz CPU, 24 GB memory, 100GB disk 4 with 24-core 2.2 GHz CPU, 32 GB memory, 500 GB disk |
| Network | ... |
- Python 3.6
- numpy == 1.21.5
- pandas == 1.4.4
- torch == 1.13.1
pip3 install -r requirements.txt
Docker 20.10.12
A running Kubernetes cluster is required before deploying DeepScaler. The following instructions are tested with Kubernetes v1.23.4. For set-up instructions, refer to this.
Prometheus is an open-source monitoring and alerting toolkit used for collecting and storing metrics from various systems. For detailed installation steps, please refer to this.
Istio is an open-source service mesh platform that enhances the management and security of microservices in a distributed application. After having a cluster running a supported version of Kubernetes, installing Istio is needed. Follow these steps to get started with Istio.
We utilize the Locust load testing tool, an open-source tool that employs Python code to define user behaviors and simulate millions of users.
We use Jaeger to initiate scenario A, which involves establishing an invocation relationship among microservices through automated processes. Follow these steps to get started with Jaeger.
(1) kubectl create -f <(istioctl kube-inject -f /benchmarks/bookinfo/bookinfo.yaml)
(2) kubectl apply -f /benchmarks/bookinfo/bookinfo-gateway.yaml)
(1) kubectl create -f <(istioctl kube-inject -f /benchmarks/boutique/boutique.yaml)
(2) kubectl apply -f /benchmarks/boutique/boutique-gateway.yaml)
Deploy the Train-Ticket system on K8S with istio.
(1) kubectl create -f <(istioctl kube-inject -f /benchmarks/train-ticket/ts-deployment-part1.yml)
(2) kubectl create -f <(istioctl kube-inject -f /benchmarks/train-ticket/ts-deployment-part2.yml)
(3) kubectl create -f <(istioctl kube-inject -f /benchmarks/train-ticket/ts-deployment-part3.yml)
(4) kubectl apply -f /benchmarks/train-ticket/trainticket-gateway.yaml
The generated workload intensity varied over time, emulating typical characteristics of microservice workloads, including slight increases, slight decreases, sharp increases, sharp decreases, and continuous fluctuations. The flow data simulation script is collected from FIFA World Cup access datasets and stored in the file.
The script load_generator.py is for simulating user behavior for both "Bookinfo" and "Online-Boutique" and load_generator_train.py is for simulating user behavior for "Train-Ticket."
Simulate the workload generator:
sh sendFlow/sendLoop.sh
You can refer to this webpage for customized usage.
The information that needs to be configured before model training is stored in config/train_config.yaml, and the processed data sets and various model configuration information are stored in config/train_datasets_speed.yaml. You can modify the tuning parameters yourself.
template = {
"cpu":"sum(irate(container_cpu_usage_seconds_total{{container=~'{1}',namespace=~'{0}'}}[1m]))/sum(container_spec_cpu_quota{{container=~'{1}',namespace=~'{0}'}}/container_spec_cpu_period{{container=~'{1}',namespace=~'{0}'}})",
"mem": "sum(container_memory_usage_bytes{{namespace='{0}',container='{1}'}}) / sum(container_spec_memory_limit_bytes{{namespace='{0}',container='{1}'}})",
"res": "sum(rate(istio_request_duration_milliseconds_sum{{reporter='destination',destination_workload_namespace='{0}',destination_workload='{1}'}}[{2}]))/sum(rate(istio_request_duration_milliseconds_count{{reporter='destination',destination_workload_namespace='{0}',destination_workload='{1}'}}[{2}]))/1000",
"req": "sum(rate(istio_requests_total{{destination_workload_namespace='{0}',destination_workload='{1}'}}[{2}]))",
"pod": "count(container_spec_cpu_period{{namespace='{0}',container='{1}'}})"
}
Run this code, and the data will be stored in the folder location you have set.
python metrics_fetch.py
#Here's "Online-Boutique" example:
prefix_api = "http://localhost:30090/api/v1/query?query="
namespace = 'boutiquessj'
interval = 120
services = ["adservice", "cartservice", "checkoutservice","currencyservice","emailservice","frontend","paymentservice","productcatalogservice","recommendationservice","shippingservice"]
metrics = ['cpu','res','req','mem','pod']
Transform the raw dataset into a time-sliced dataset for model training and learning.
python data_process.py
We provide a more detailed and complete command description for training and testing the model:
python -u main.py
--model_config_path <model_config_path>
--train_config_path <train_config_path>
--model_name <model_name>
--num_epoch <num_epoch>
--num_iter <num_iter>
--model_save_path <model_save_path>
--max_graph_num <max_graph_num>
| Parameter name | Description of parameter |
|---|---|
| model_config_path | Config path of models |
| train_config_path | Config path of Trainer |
| model_name | Model name to train |
| num_epoch | Training times per epoch |
| num_iter | Maximum value for iteration |
| model_save_path | Model save path |
| max_graph_num | Volume of adjacency matrix set |
More parameter information please refer to main.py. The models exported after running the file are stored in the model_states.
Utilizing well-trained and tested models to enable automatic scaling of various microservices.
python predict_scale.py
python similarity.py
python calculate.py
If you find this repository useful in your research, please consider citing the following papers:
@INPROCEEDINGS{10298341,
author={Meng, Chunyang and Song, Shijie and Tong, Haogang and Pan, Maolin and Yu, Yang},
booktitle={2023 38th IEEE/ACM International Conference on Automated Software Engineering (ASE)},
title={DeepScaler: Holistic Autoscaling for Microservices Based on Spatiotemporal GNN with Adaptive Graph Learning},
year={2023},
pages={53-65},
keywords={Learning systems;Cloud computing;Costs;Fluctuations;Microservice architectures;Feature extraction;Dynamic scheduling;Cloud Computing;Microservice;QoS;Resource Management;Holistic Autoscaling;Graph Convolution;Container},
doi={10.1109/ASE56229.2023.00038}}
If you have any questions, feel free to contact Shijie Song or Chunyang Meng through Email ([email protected], [email protected]) or Github issues. Pull requests are highly welcomed!