e-commerce search benchmark

Introduction

E-commerce search benchmark is the first end-to-end application benchmark for e-commerce search system with personalized recommendations. It helps people deeply understand the characteristics of workloads for e-commerce search and drives better design options for industry search system.

Features of the benchmark:

Provides a Data Generator using real-world datasets and producing synthetic data of various scales.
Provides a Workload Generator that driven by the real-world user logs from Taobao.
Provides a e-commerce search model eSearchEngineModel that simulated the search system of Taobao.
Evaluates the overall performance and performance of individual components.

The e-commerce search benchmark is built on docker images. As shown in the figure above,the benchmark consists of 7 docker images:

aliesearch-search-planner
aliesearch-query-planner
aliesearch-tf-serving
aliesearch-ha3
aliesearch-ranking-service
aliesearch-jmeter-image
aliesearch-benchmark-cli

where, images 1~5 constitute the e-commerce search model eSearchEngineModel, aliesearch-benchmark-cli is the Data Generator, aliesearch-jmeter-image drives the workload generated by Workload Generator to the e-commerce search model eSearchEngineModel.

Preparation

Dependency

The following build tools are required：

gradle 4.x
maven 3.x
docker 17.09+
jdk8

Note: Make sure you can use docker without sudo by running

sudo usermod -aG docker $USER

Build

Run build.sh in eSearchEngineModel directory to build and publish image, etc.

Compile and build docker images
```
./build.sh build
```
Publish the docker images to specified remote docker reposition. If you are only running locally, you can skip this step.
```
./build.sh push ${repo_name}
```

Running benchmark

Note:

Correctly value for java maximum memory size must be required, otherwise the docker images won't work. You can set it for temporary by running

sudo sysctl -w vm.max_map_count=262144

or you can set it for permanent by directly edit the /etc/sysctl.conf file on the host, adding a line as flollow:

vm.max_map_count = 262144

and then run

sysctl -p

Running Example in standalone

Copy appctl.sh in this directory to your working directory;
Pull all the images by running: (If you are only running locally, do skip this step.)
```
 sudo ./appctl.sh pull
```
Start all images and automatically import the default data by running:
```
 sudo ./appctl.sh start
```

Check if all the images are ready to work by running:

 curl -H 'Content-Type:application/json;charset=UTF-8' -d'
             {"uid":"798", "page":0, "query":"68"}' 172.17.0.6:8080/search

Run an example experiment by jmete:

sudo docker exec -it aliesearch-jmeter-image bash

run to start the pressure test process by running:

cd apache-jmeter-5.1.1
./bin/jmeter -n -t search_stress.jmx -l result -e -o report

Running Example on k8s

Distributed deployment is also provided depending on k8s, which is more in line with the online environment. The benchmark can be run on a k8s cluster as follows:

Copy k8s.yml to working directory on k8s master node.
Replace the docker images repo configuration ink8s.yml with your own repo
Pull all images and automatically import the default data by running:
```
sudo kubectl create -f k8s.yml
```
Check the pods status to ensure that all pods start up properly by running:
```
sudo kubectl get pod
```

Run an example experiment by jmeter:

sudo kubectl exec -it aliesearch-jmeter-image bash

run to start the pressure test process

cd apache-jmeter-5.1.1
./bin/jmeter -n -t search_stress.jmx -l result -e -o report

Running with Customized Pattern

The benchmark comes with an e-commerce data generator and workloads generator that driven by production data and real-world user queries. There are 4 steps for running an experiment with Customized data scale and workload pattern.

1. Install the benchmark

Start all the images of the benchmark according to the 1~3 steps in Running Example in standalone

2. Data Generation

sudo docker exec -ti aliesearch-benchmark-cli bash

Generate goods and user data , load them into corresponding search components by running:

vim entrypoint.sh
sh entrypoint.sh ${scale_factor}

where, ${scale_factor} sets the scale factor determining the dataset size (1 scale factor equals 1G goods and 6K user, 10 scale factor equals 10G goods and 60K user,and so on)

3. Workload Generation

Change to the directory workload_generator, and generate workload for specified hour of one day and specified user scale by running:

python3 workload_generator -t 21 -u 100000

A query_workload.csv file will be generated in the directory workload_generator, which will be driven to system model under test (e.g. eSearchEngineModel ).

4. Run experiment by jmeter

Firstly, Copy the generated workload file query_workload.csv to jmeter path (apache-jmeter-5.1.1/bin/jmeter) in jmeter docker. Then, login to 'jmeter-image' by running:

sudo docker exec -ti aliesearch-jmeter-image bash

And change to the jmeter bin directory, start the pressure test process by running:

cd apache-jmeter-5.1.1
./bin/jmeter -n -t search_stress.jmx -l result -e -o report

Benchmark Results:Example Output

1. End-to-end performance metrics

E-commerce search benchmark use Jmeter as the driver to drive workload described query_workload.csv to system under test, so the end-to-end performance metrics are displayed by CLI and HTML of Jmeter. An example CLI of Jmeter CLI report is as following:

l result -e -o repo2rtpache-jmeter-5.1.1$ ./bin/jmeter -n -t search_stress.jmx -l Creating summariser
Created the tree successfully using search_stress.jmx Starting the test @ Wed Nov 13 02:33:53 UTC 2019 (1573612433454) Waiting for possible Shutdown/StopTestNow/HeapDump/ThreadDump message on port 4445

summary + 67 in 00:00:06 = 10.7/s Avg: 45 Min: 15 Max: 115 Err: 0 (0.00%) Active: 3 Started: 3 Finished: 0 summary + 3108 in 00:00:30 = 103.6/s Avg: 40 Min: 13 Max: 97 Err: 0 (0.00%) Active: 8 Started: 8 Finished: 0 summary = 3175 in 00:00:36 = 87.5/s Avg: 40 Min: 13 Max: 115 Err: 0 (0.00%)
summary + 6645 in 00:00:30 = 221.5/s Avg: 40 Min: 12 Max: 134 Err: 0 (0.00%) Active: 10 Started: 10 Finished: 0

A detail result is reported by HTML in the report direcoty specified in runnign experiment. An example HTML report of Jmeter is as following:

2. Response time breakdwon

All steps of the full search path are logged, the statistical analysis of the breakdown of response time is printed every minute. To check the statistics of breakdown time, login to 'aliesearch-search-planner' by running:

sudo docker exec -it aliesearch-search-planner bash

Then check the statistics in out.log by running:

vim out.log

and the statistics of breakdown time is recorded as following:

SearchBench specifies an e-commerce search dataset and workloads that driven by production data and real-world user queries, and target at the hardware and software system that provide e-commerce online search service. Therefore, any of those system can be used to establish the feasibility of this benchmark.

chen0031 / eCommerceSearchBench