YCSB/rest

Quick Start

This section describes how to run YCSB to benchmark HTTP RESTful webservices. The aim of the rest binding is to benchmark the performance of any sepecific HTTP RESTful webservices with real life (production) dataset. This must not be confused with benchmarking various webservers (like Apache Tomcat, Nginx, Jetty) using a dummy dataset.

1. Set Up YCSB

Clone the YCSB git repository and compile:

git clone git://github.com/brianfrankcooper/YCSB.git
cd YCSB
mvn -pl site.ycsb:rest-binding -am clean package

2. Set Up an HTTP Web Service

There must be a running HTTP RESTful webservice accesible from the instance on which YCSB is running. If the webservice is running on the local instance default HTTP port 80, it's base URL will look like http://127.0.0.1:80/{service_endpoint}. The rest binding assumes that the webservice to be benchmarked already has a valid dataset. THe rest module has been designed in this way for two reasons:

1. The performance of most webservices depends on the size, pattern and the nature of the real life dataset accesible from these services. Hence creating a dummy dataset might not actually reflect the true performance of a webservice to be benchmarked.

2. Since many webservices have a non-naive backend which includes interaction with multiple backend components, tables and databases. Generating a dummy dataset for such webservices is a non-trivial and a time consuming task.

However to benchmark a webservice before it has access to a real dataset, support for automatic data insertion can be added in the future. An example of such a scenario is benchmarking a webservice before it moves to production.

3. Run YCSB

At this point we assume that you've setup a webservice accesible at an HTTP endpoint like this: http://{host}:{port}/{service_endpoint}.

Before you are ready to run please ensure that you have prepared a trace for the CRUD operations to benchmark your webservice.

Trace is a collection of URL resources that should be hit in order to benchmark any webservice. The more realistic this collection of URL is, the more reliable and accurate are the benchmarking results because this means simulating the real life workload more accurately. Tracefile is a file that holds the trace. For example, if your webservice exists at http://{host}:{port}/{endpoint}, and you want to benchmark the performance of READS on this webservice with five resources (namely resource_1, resource_2 ... resource_5) then the url.trace.read file will look like this:

http://{host}:{port}/{endpoint}/resource_1 http://{host}:{port}/{endpoint}/resource_2 http://{host}:{port}/{endpoint}/resource_3 http://{host}:{port}/{endpoint}/resource_4 http://{host}:{port}/{endpoint}/resource_5

The rest module will pick up URLs from the above file according to the requestdistribution property (default is zipfian) mentioned in the rest_workload. In the example above we assume that the property url.prefix (see below for property description) is set to empty. If url.prefix property is set to http://{host}:{port}/{endpoint}/ the equivalent of the read trace given above would look like:

resource_1 resource_2 resource_3 resource_4 resource_5

In real life the traces for various CRUD operations are diffent from one another. HTTP GET will rarely have the same URL access pattern as that of HTTP POST or HTTP PUT. Hence to give enough flexibility to benchmark webservices, different trace files can be used for different CRUD operations. However if you wish to use the same trace for all these operations, just pass the same file to all these properties - url.trace.read, url.trace.insert, url.trace.update & url.trace.delete.

Now you are ready to run! Run the rest_workload:

./bin/ycsb run rest -s -P workloads/rest_workload

For further configuration see below:

Default Configuration Parameters

The default settings for the rest binding are as follows:

• url.prefix

  • The base endpoint URL where the webservice is running. URLs from trace files (DELETE, GET, POST, PUT) will be prefixed with this value before making an HTTP request. A common usage value would be http://127.0.0.1:8080/{yourService}
  • Default value is http://127.0.0.1:8080/.

• url.trace.read

  • The path to a trace file that holds the URLs to be invoked for HTTP GET method. URLs must be seperated by a newline.

• url.trace.insert

  • The path to a trace file that holds the URLs to be invoked for HTTP POST method. URLs must be seperated by a newline.

• url.trace.update

  • The path to a trace file that holds the URLs to be invoked for HTTP PUT method. URLs must be seperated by a newline.

• url.trace.delete

  • The path to a trace file that holds the URLs to be invoked for HTTP DELETE method. URLs must be seperated by a newline.

• headers

  • The HTTP request headers used for all requests. Headers must be separated by space as a delimiter.
  • Default value is Accept */* Accept-Language en-US,en;q=0.5 Content-Type application/x-www-form-urlencoded user-agent Mozilla/5.0

• timeout.con

  • The HTTP connection timeout in seconds. The response will be considered as an error if the client fails to connect with the server within this time limit.
  • Default value is 10 seconds.

• timeout.read

  • The HTTP read timeout in seconds. The response will be considered as an error if the client fails to read from the server within this time limit.
  • Default value is 10 seconds.

• timeout.exec

  • The time within which request must return a response. The response will be considered as an error if the client fails to complete the request within this time limit.
  • Default value is 10 seconds.

• log.enable

  • A Boolean value to enable console status logs. When true, it will print all the HTTP requests being made and thier response status on the YCSB console window.
  • Default value is false.

• readrecordcount

  • An integer value that signifies the top k URLs (entries) to be picked from the url.trace.read file for making HTTP GET requests. Must have a value greater than 0. If this value exceeds the number of entries present in url.trace.read file, then k will be set to the number of entries in the file.
  • Default value is 10000.

• insertrecordcount

  • An integer value that signifies the top k URLs to be picked from the url.trace.insert file for making HTTP POST requests. Must have a value greater than 0. If this value exceeds the number of entries present in url.trace.insert file, then k will be set to the number of entries in the file.
  • Default value is 5000.

• deleterecordcount

  • An integer value that signifies the top k URLs to be picked from the url.trace.delete file for making HTTP DELETE requests. Must have a value greater than 0. If this value exceeds the number of entries present in url.trace.delete file, then k will be set to the number of entries in the file.
  • Default value is 1000.

• updaterecordcount

  • An integer value that signifies the top k URLs to be picked from the url.trace.update file for making HTTP PUT requests. Must have a value greater than 0. If this value exceeds the number of entries present in url.trace.update file, then k will be set to the number of entries in the file.
  • Default value is 1000.

• readzipfconstant

  • An double value of the Zipf's constant to be used for insert requests. Applicable only if the requestdistribution = zipfian.
  • Default value is 0.9.

• insertzipfconstant

  • An double value of the Zipf's constant to be used for insert requests. Applicable only if the requestdistribution = zipfian.
  • Default value is 0.9.

• updatezipfconstant

  • An double value of the Zipf's constant to be used for insert requests. Applicable only if the requestdistribution = zipfian.
  • Default value is 0.9.

• deletezipfconstant

  • An double value of the Zipf's constant to be used for insert requests. Applicable only if the requestdistribution = zipfian.
  • Default value is 0.9.