/sinan-local

Primary LanguageC

SINAN-LOCAL

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

  • Ubuntu 18.04
  • Docker 19.03
  • Docker swarm latest
  • Python 3.5+
  • MXNet
  • XGBoost

Code structure

Similar to https://github.com/zyqCSL/sinan-gcp

Usage

Generating cluster configurations

Please make sure to generate the configuration files specific to your local cluster. The config directory in this repo contains an example configuration with 2 serveres each with 88 cpus, and a dedicated gpu server for inference. In order to reproduce experiments in the paper, cpu resoruces in the cluster are expected to be no less than the provided configuration.

Service cluster configuration (docker_swarm/misc/make_cluster_config.py)

For exmaple, in docker_swarm/misc, python3 make_cluster_config.py --nodes ath-8 ath-9 --cluster-config test_cluster.json --replica-cpus 4 generates the cluster configuration for deploying SocialNetwork using two servers ath-8 and ath-9, and saves it as docker_swarm/config/test_cluster.json.

For adapting to your own local cluster, please check the following:

SocialNetwork
HotelReservation

Inference engine configuration (docker_swarm/misc/make_gpu_config.py)

In docker_swarm/misc, python3 make_gpu_config.py --gpu-config gpu.json generates the predictor configuration for SocialNetwork and saves it in docker_swarm/config/gpu.json. Similarly, python3 make_gpu_hotel_config.py --gpu-config gpu_hotel.json generates predictor configuration for hotel reservation.

For SocialNetwork, please modify https://github.com/zyqCSL/sinan-local/blob/master/docker_swarm/misc/make_gpu_config.py#L31-L34 to adapt to your own cluster

  • gpu_config['gpus'] to the list of gpus to use, [0] if you only have one
  • gpu_config['host'] to the ip of your own gpu server
  • gpu_config['working_dir'] to your own working directory of predictor (the path of ml_docker_swarm of the cloned repo).

Modifications are similar for HotelServation.

Setting up docker swarm cluster

The following steps assume you are in docker_swarm.

On the master node, in docker_swarm directory, execute python3 setup_swarm_cluster.py --user-name USERNAME --deploy-config test_cluster.json, in which test_cluster.json is the generated cluster configuration. You can use docker node ls to make sure that the required servers are addeded and docker node inspect SERVERNAME to inspect they are tagged properly (check "Spec"::"Label" in the output json).

This step is not necessary for HotelReservation, since its deployment is controlled by separate docker-compose files.

Data collection

Short cut scripts for data collection & deployment can be found in docker_swarm/scripts. For example, within docker_swarm/scripts, executing run_data_collect.sh will collect training samples of SocialNetwork for concurrent user number from 2 to 48. Generated data are stored in docker_swarm/logs/collected_data.

For adapting to your own local cluster, please modify:

  • --deploy-config to your own cluster configuration (generated before)
  • --user-name Your own username
  • --exp-time the running time for each user number
  • --min-users, --max-users, --users-step. Generated training will include [min_users, max_users, users_step]. If you are using a smaller cluster, you might want to scale down the max_users. This repo uses 2 88-vcpu servers, and you can scale down the max-users proportionally, or you can start from short experiments and check when your cluster saturates.

Similarly, run_data_collect_hotel.sh collects data for HotelReservation, and modifications are similar to SocialNetwork, to adapt to your own cluster. Generated data are saved in docker_swarm/logs/hotel_collected_data.

Modeling training

GPU is required for model training. The following steps assume you are in ml_docker_swarm. We also provide models trained on our own cluster, saved in ml_docker_swarm/xgb_model and ml_docker_swarm/model.

SocialNetwork

  • Execute python3 data_parser_socialml_next_k.py --log-dir LOGDIR --save-dir DATADIR to format the training dataset. LOGDIR is the directory of the raw training set (ocker_swarm/logs/collected_data), and DATADIR points to the path to save the formatted training samples. Please record the first dimension of output glob_sys_data_train.shape, which is your training data set size and you will need it later.

  • Execute python train_cnvnet.py --num-examples NUMSAMPLES --lr 0.001 --gpus 0,1 --data-dir DATADIR --wd 0.001 to train the CNN. NUMSAMPLES is your training dataset size. If you only have 1 gpu, please change to --gpus 0. Generated models are saved in ml_docker_swarm/model. You can also read the accuracy from the output log, whose name is test_single_qps_upsample by default.

  • Execute python xgb_train_latent.py --gpus 0,1 --data-dir DATADIR to train the XGBoost model. Generated models are saved in ml_docker_swarm/xgb_model.

HotelReservation

Instructions are similar to SocialNetwork, including the following steps:

  • Execute python3 data_parser_hotel_next_k.py --log-dir LOGDIR --save-dir DATADIR to format the training dataset

  • Execute python train_hotel_cnvnet.py --num-examples NUMSAMPLES --lr 0.001 --gpus 0,1 --data-dir DATADIR --wd 0.001 to train the CNN.

  • Execute python xgb_train_hotel_latent.py --gpus 0,1 --data-dir DATADIR to train the XGBoost model. Generated models are saved in ml_docker_swarm/xgb_model.

Deployment

GPU is required for deployment. The shortcut scripts for deployment are in docker_swarm/scripts.

SocialNetwork

Static load

test_deploy.sh tests the deployment situation with concurrent users in [5, 45, 5]. Execution logs are saved in docker_swarm/logs/deploy_data.

To adapt to your own cluster, please modify the following:

  • Add --deploy flag if the application is not already deployed.

  • --slave-port & --gpus-port are the ports master connects to slave and gpu server, please make sure there's no conflict.

  • --deploy-config & --gpu-config should be changed to your own configuration previously generated

  • --min-users, --max-users, --users-step should be scaled proportionally to your cluster size

  • --user-name to your own.

Diurnal load

test_deploy.sh tests the diurnal pattern where concurrent users start as 4, gradually rises to 36 and then gradually drops back to 4, with each period lasting 120s. The instructions are the same as static load. Execution logs are saved in docker_swarm/logs/diurnal_deploy_data.

HotelReservation

test_deploy_hotel.sh tests the constant users and test_deploy_diurnal_hotel.sh tests diurnal pattern. The modification instructions are similar to SocialNetwork. Execution logs are saved in docker_swarm/logs/hotel_deploy_data.

Data processing

For SocialNetwork, python data_proc/count_cpus.py LOGDIR calculates the average cpu usage, tail latencies & violation rates of the execution logs, and python data_proc/plot.py LOGDIR plots the real-time cpu allocation of each service and end-to-end latencies. LOGDIR should be the execution logs generated from deployment.

For HotelReservation, python data_proc/count_cpus_hotel.py LOGDIR is similar to that of SocialNetwork.