SINAN-GCP

Publication

If you use Sinan in your research, please cite our ASPLOS'21 paper.

@inproceedings{sinan-asplos2021,
author = {Yanqi, Zhang and Weizhe, Hua and Zhuangzhuang, Zhou and G. Edward, Suh and Christina, Delimitrou
},
title = {Sinan: ML-Based & QoS-Aware Resource Management for Cloud Microservices},
booktitle = {Proceedings of the Twenty-Sixth International Conference on Architectural Support for Programming Languages and Operating Systems},
series = {ASPLOS '21}
}

Prerequisites

• Python 3.5+
• Python 2.7 (for plotting)
• Install & set up Google Cloud SDK (https://cloud.google.com/sdk/docs/how-to). In order to reproduce the results presented in the paper, the CPU quota (Compute Engine API) of your Google Cloud project should be no less than 500.

Code structure

benchmarks

benchmarks directory contain the source codes of tested benchmarks.

For SocialNetwork application (benchmarks/socialNetwork-ml-swarm), we added two compute-intensive machine learning microservices (text-filter and media-filter), and also add image data to user posts (previously posts only include text), in order to make the application a little more representative than original versions. We also provide a warm up script (benchmarks/socialNetwork-ml-swarm/setup_social_graph_init_data_sync.py) for the application, to fill the social-network friendship graph, and to fill the databases with posts

config

configuration files of cluster, scheduling actions, and inference engine

misc

scripts to generate config files

exp

scripts for running experiments

locust

code for workload generation

ml

ml models and scirpts for data preparation, training, deployment and fine-tunning (for different user workload patterns). The complete flow includes the following steps:

• collect training data (short cut script in exp_scripts/gcloud_run_exp.sh)
• process collected data with data_parser_docker_next_k.py
• train the CNN & XGBoost model with processed data (train_cnvnet.py & xgb_trian_latent.py). The architectures of neural networks are in the model directory.
• deploy online with running microservices (social_media_predictor.py)
• fine tune the model to adapt to changes, cluster changes, workload skews e.g. (finetune.sh)

src

utilization functions

scripts

initialization scripts for GCE VMs

root directory

master_data_collect_social.py --- master for data collection

master_deploy_social.py --- master for running deployment experiment of the social network benchmark

master_deploy_diurnal_social.py --- master for running deployment experiment of the social network benchmark with diurnal request per second (rps) pattern

slave_data_collect.py --- slave for data collection & deployment

gcloud.py --- set up gcloud cluster & collect data

gcloud_deploy.py --- set up glcoud cluster & deploy the social network benchmark

gcloud_deploy_diurnal.py --- set up gcloud cluster & deploy the social network benchmark with diurnal rps pattern

Usage

Following instructions assume that users start from git root directory. Before executing any shell script, users should make sure to clone the repo to their home directory and change the '--username' argument in the shell script to his own Google Cloud user name. When execution of scripts is completed, system execution log should be in the logs directory of the master node. Users can copy the data to local machine with scp (the ssh keys are automatically generated and stored in keys directory)

Deployment experiment with static rps and identical workload

This script tests the deployment of Sinan under static RPS, with the workload composition the same as training data (w0 in Figure 13 and Figure 14 in the paper). For detailed information on workload characterization, please check locust/src/social_rps_10.py

cd exp_scripts
./gcloud_deploy_cpu_pred.sh

Deployment experiment with static rps and skewed workload

This script tests the deployment of Sinan under static rps, with the workload composition slightly skewed from training data (w1-w3 in Figure 13 and Figure 14 in the paper). For detailed information on workload characterization, please check locust/src/social_rps_10_v*.py

cd exp_scripts
./gcloud_deploy_cpu_pred_skewed.sh x

x can be set to 1, 2 and 3.

Deployment experiment with diurnal rps pattern

This script tests the deployment of Sinan under a diurnal rps pattern

cd exp_scripts
./gcloud_deploy_diurnal_cpu_pred.sh

Data processing

python data_proc/count_cpu.py dir_name

This script calculates the average cpu usage, tail latencies & violation rates of the execution logs.

python data_proc/plot.py dir_name

This script plots the real-time cpu allocation of each service and end-to-end latencies.

dir_name should be the directory that contains system execution log. For example, logs/deploy_data in the static load experiments. For specific log directory, check DataDir variable in the master_deploy* scripts.