/wsdemo

A Cowboy Websocket demo

Primary LanguageErlangOtherNOASSERTION

wsdemo

A Websocket competition. Will your favorite platform come out on top? If you think your platform of choice was poorly represented, "submit a pull request™"

How do you include a contestant?

Just create the "Hello, World!" of socket servers, an echo server that speaks WebSocket version 13 (13 is the only version my client supports).

First write your implementation in place it in the competition/ directory.

Added any instructions needed to install dependencies in configure_ubuntu-12.04.sh. Add any build instructions to competition/Makefile if your project needs to be compiled

wsdemo_bench utilizes supervisord to start and stop the servers while running the tests. In order to include your server in the test you need to create a supervisord ini file in competition/servers.

[program:%(server_name)s]
command=%(command)s
autostart=false
autorestart=false
startretries=0
stopasgroup=true
killasgroup=true

The server name must follow the following format:

{language}-{platform}

For instance if you wrote an echo server using bash and nc the server name would be bash-nc. You can add additional demarcations if needed. For instance the Python tornado example has a single threaded version and a multiprocessor version which go by the name python-tornado-1 and python-tornado-N.

Next, you need to add your server to the server list in wsdemo_bench.app.src configuration file.

Finally, add yourself to the AUTHORS file; you've earned it.

Running the benchmark

It is best to run the benchmark client on a separate machine than the servers.

Using two brand new Ubuntu 12.04 servers, run the configure_ubuntu-12.04.sh script on each. This script will destructively alter each server so use caution and review the script before running it.

If you want to use a different OS, use configure_ubuntu-12.04.sh as a template for configuring your system of choice.

Next compile the code on both machines:

  make

To run a single benchmark on a server do the following:

./runtest db_path duration hostname port client_count

For instance for a benchmark of 192.168.1.5:8000 of 1000 clients lasting 60 seconds with with the data stored at data/myserver the command would be:

./runtest data/myserver 60 192.168.1.5 8000 1000

Running the entire suite on your own servers

There are two components in wsdemo_bench. The first component is supervisord

wsdemo_bench communicates with Supervisord to start and stop each server before each benchmark.

On the machine that you are running the servers on do the following:

sudo bash
ulimit -n 999999
cd competition
supervisord

You can monitor supervisord by using the supervisorctl command:

competition/ $ supervisorctl status

Next, on the client machine create a config/small-scale.config file:

[
    {sasl, [
        {sasl_error_logger, {file, LogFile}}]},

    {wsdemo_bench, [
                {host, Ip},
                {port, 8000},
                {db_root, DbRoot},
                {clients, 100},
                {seconds, 10},
                {supervisord, {ServerHost, 9001}}]}].

Replace LogFile, Ip, DbRoot and ServerHost with the correct values. Ensure that both the path to LogFile and the path to DbRoot exist or the test will crash.

This configuration file describe a test using 100 clients for 10 seconds each test. We will use this smaller test to ensure that all the servers start and stop properly.

To run this benchmark, do the following:

sudo bash
ulimit -n 999999
./bin/run_all_tests.sh -config config/small-scale.config

If all goes well, you should see the suite running in front of your eyes. Let that run for the entire test. It should run for 12 minutes or so.

If any of the servers crashes you can do the following to diagnose the cause:

competition/ $ supervisorctl tail $SERVER_NAME stderr

After the small scale test, it is time to run the full scale test:

[
    {sasl, [
        {sasl_error_logger, {file, LogFile}}]},

    {wsdemo_bench, [
                {host, Ip},
                {port, 8000},
                {db_root, DbRoot},
                {supervisord, {ServerHost, 9001}}]}].

Suite config

Here is the full spec for the wsdemo_bench config:

 [
    {host, Host :: string()},
    {port, Port :: integer()},
    {db_root, DbRoot :: directory()},
    {supervisord, {Host, 9001}},
    {clients, number()}, % defaults to 10000
    {seconds, number()}, % defaults to 600
    {servers, [ServerName1::string(), ..., ServerNameN::string()]} % defaults to full suite
 ]        

Exporting the data

The resulting leveldb databases will be placed in your configured db_root. The events are stored as binary Erlang terms in the database so you will need to export the events to use them.

There are two scripts to do that. ./bin/compile_all_stats.sh and ./bin/convert_all_stats.sh

./bin/compile_all_stats.sh creates three .csv tables:

  • counts.csv - the sums of all the events
  • handshake_times.csv - timestamp, elapsed usecs for each handshake
  • message_latencies.csv - timestamp, elapsed usecs for each message

./bin/convert_all_stats.sh dumps the raw events as a events.csv table.

The events data has the following fields:

timestamp, type, client_id, event_key, event_data
  • timestamp - The timestamp of the event
  • type - the server type
  • client_id - A string that identifies the client
  • event_key - The event type
  • event_data - event specific data

The event_data is used to pair ws_init and ws_onopen events to calculate the elapsed handshake time and the event_data is used to pair send_message events to recv_message events to calculate the elapsed message time. The event_data for 'EXIT' events take the format of {pid(), Reason :: any()}. These are some of the expected reasons for a client crash:

connection_timeout - The client could not establish a connection
                     with the server in under 2 seconds
normal - The server closed the connection.  This should never
         happen for this benchmark

Any reason other than these are an unexpected error and should be reported as an issue. If you see a {error, Reason} you should probably file a bug report as well unless you can determine it is a server issue.