EastTransitTenants

Introduction

This project is a performance profiling and bottleneck analysis framework built on top of an open-source railway ticketingmicroservices system. From a big picture,our framework consists of five components:

• orchestrator:coordinate the interaction of other functional components;
  
  
• load generator: generate workload for training and profiling purpose;
  
  
• benchmark application: an opensource railway ticketing microservices system which contains 41 mi-croservices and whose request can involve up to 200 spans (You can get more details at https://github.com/FudanSELab/train-ticket/wiki);
  
  
• bottleneck discover: find all potential bottleneck services ina time-efficient way;
  
  
• performance profiler: instrumentallyprofile how a service influencing the end-to-end performanceof the requests it belongs to.

Architecture Graph

Dependencies

• Golang
• python dependencies
  • pandas
  • sklearn
  • joblib

Usage

First, build the framework with: go build

After built-up, you can choose either to train the bottleneck-discovery model or start the performance profiling and bottleneck analysis by run the project with different flags.

Train Bottleneck-discovery Model

1. Configuration for training

The input configurations are defined in train_config.json . We have provided a template configuration in ths file. You can modify it based on your application and request types.

{
    "request": [{
        "name": "search_tickets",
        "url": "http://35.231.88.215:32677/api/v1/travel2service/trips/left",
        "body": "{\"startingPlace\": \"Shang Hai\",\"endPlace\": \"Tai Yuan\",\"departureTime\": \"2020-12-21\"}"
    }],
    "bearer": "",
    "jaeger_ip": "35.231.88.215:32688",
    "workload": [1, 10, 25, 50, 100],
    "target_serv" : "ts-ui-dashboard.default"
}

• request: list of requests to generate training data. Each of them should contain the name of the request, the url of its API call, and the request body in json format
• bearer: authorization bearer token. Only provided if needed
• jaeger_ip: the IP address of Jaeger
• target_serv: the entrance service for all requests to your application, which is normally the front-end service.

2. Train the model

go build
./easttransittenants -type=train

We provided some example training data under /data directory

Profile the Bottleneck

1. Configuration for profiling

The configurations for profiling are defined in profile_config.json.

{
    "jaeger_ip": "35.231.88.215:32688",
    "request": {
            "bearer": "",
            "name": "search_tickets",
            "url": "http://35.231.88.215:32677/api/v1/travel2service/trips/left",
            "body": "{\"startingPlace\": \"Shang Hai\",\"endPlace\": \"Tai Yuan\",\"departureTime\": \"2020-12-21\"}"
    },
    "workload": 10,
    "target_serv": "ts-travel2-service",
    "precision_ms": 25
}

• jaeger_ip: the address of the Jaeger deployment, used by our jaeger client to query trace data
• request: the target request to profile
• workload: target workload level
• target_serv: the entrance service for the request to the application
• precision_ms: desired precision of the bottleneck switching boundary in milliseconds

2. Do Profile

go build
./easttransittenants -type=profile