ruby-mqtt 2x slower than paho-mqtt #115

rgaufman · 2018-05-10T16:01:12Z

Hi there,

I'm comparing different ruby mqtt clients, I have this code:

client = MQTT::Client.new(username: 'testuser', password: 'testpasswd', client_id: "client_#{rand(1..1_000_000)}", host: '127.0.0.1', )
client.connect()

count = 0
loop do
  count += 1
  i = rand(1..10)

  topic = "to/timebox#{i}/cameras"
  msg = "message #{count} #{Time.now} for timebox#{i}"
  client.publish(topic, msg, true, 1)
end

Here is a version using paho-mqtt

$client = PahoMqtt::Client.new(username: 'testuser', password: 'testpasswd', client_id: "client_#{rand(1..1_000_000)}")
$client.connect('127.0.0.1', 1883)

$count = 0
def send_message
  $count += 1
  i = rand(1..$total)

  topic = "to/timebox#{i}/cameras"
  msg = "message #{$count} #{Time.now} for timebox#{i}"
  $client.publish(topic, msg, true, 1)
end

$client.on_puback { send_message } # QOS1
send_message
loop { sleep 1 } # to keep process running

I used pv to calculate how many messages are being received per second and I'm getting:

ruby-mqtt: 92 / sec
paho-mqtt: 180 /sec

I prefer the syntax of ruby-mqtt and it appears to be better maintained, but any ideas why ruby-mqtt is nearly 2x slower?

njh · 2018-05-11T12:32:07Z

No sorry, I am not certain why there is such a different in publish speed - it must have something to do with ruby-mqtt being synchronous/using threads. paho-mqtt borrowed the packet parsing code from ruby-mqtt, so it is unlikely to be that.

This commit changes the `publish` method to use a queue to wait for Puback packets rather than polling a hash. Every time the read loop gets data or a timeout from `IO.select`, it will send a message to everyone waiting for a `Puback` packet. If the we're within the deadline, then the loop executes again, if we got a packet, we'll return the packet, and if we're outside the deadline, a `-1` is returned. This upside is that this patch speeds up the publish method by over 100x. Here is the benchmark: ```ruby require "securerandom" require "mqtt" require "benchmark/ips" require "stackprof" client = MQTT::Client.new(username: 'testuser', password: 'testpasswd', client_id: "client_#{SecureRandom.hex(10)}", host: '127.0.0.1') client.connect Benchmark.ips do |x| x.report("send message") { i = rand(1..10) topic = "to/timebox#{i}/cameras" msg = "message #{Time.now} for timebox#{i}" client.publish(topic, msg, true, 1) } end ``` Before this patch: ``` $ ruby -I lib thing.rb Warming up -------------------------------------- send message 8.000 i/100ms Calculating ------------------------------------- send message 85.042 (± 4.7%) i/s - 432.000 in 5.089261s ``` After this patch: ``` $ ruby -I lib thing.rb Warming up -------------------------------------- send message 915.000 i/100ms Calculating ------------------------------------- send message 9.453k (± 4.5%) i/s - 47.580k in 5.043716s ``` The downside is that the timeout isn't exact. Since `IO.select` times out every `0.5` seconds (according to the `SELECT_TIMEOUT` constant), the deadline in the `publish` method could be missed by that amount of time. Refs njh#115

This commit changes the `publish` method to use a queue to wait for Puback packets rather than polling a hash. Every time the read loop gets data or a timeout from `IO.select`, it will send a message to everyone waiting for a `Puback` packet. If the we're within the deadline, then the loop executes again, if we got a packet, we'll return the packet, and if we're outside the deadline, a `-1` is returned. This upside is that this patch speeds up the publish method by over 100x. Here is the benchmark: ```ruby require "securerandom" require "mqtt" require "benchmark/ips" require "stackprof" client = MQTT::Client.new(username: 'testuser', password: 'testpasswd', client_id: "client_#{SecureRandom.hex(10)}", host: '127.0.0.1') client.connect Benchmark.ips do |x| x.report("send message") { i = rand(1..10) topic = "to/timebox#{i}/cameras" msg = "message #{Time.now} for timebox#{i}" client.publish(topic, msg, true, 1) } end ``` Before this patch: ``` $ ruby -I lib thing.rb Warming up -------------------------------------- send message 8.000 i/100ms Calculating ------------------------------------- send message 85.042 (± 4.7%) i/s - 432.000 in 5.089261s ``` After this patch: ``` $ ruby -I lib thing.rb Warming up -------------------------------------- send message 915.000 i/100ms Calculating ------------------------------------- send message 9.453k (± 4.5%) i/s - 47.580k in 5.043716s ``` The downside is that the timeout isn't exact. Since `IO.select` times out every `0.5` seconds (according to the `SELECT_TIMEOUT` constant), the deadline in the `publish` method could be missed by that amount of time. Refs #115

njh · 2020-03-09T18:06:39Z

@rgaufman please can you test to see if @tenderlove's improvements fix your performance problems?

dentarg · 2024-06-30T16:29:54Z

Maybe just close this issue seeing https://github.com/RubyDevInc/paho.mqtt.ruby hasn't had any updates in main since Oct 4, 2018?

rgaufman changed the title ~~Why is ruby-mqtt much slower than paho-mqtt?~~ ruby-mqtt 2x slower than paho-mqtt May 10, 2018

tenderlove mentioned this issue Jan 27, 2020

Use a queue to wait for Puback packets rather than polling #120

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

ruby-mqtt 2x slower than paho-mqtt #115

ruby-mqtt 2x slower than paho-mqtt #115

rgaufman commented May 10, 2018 •

edited

Loading

njh commented May 11, 2018 •

edited

Loading

njh commented Mar 9, 2020

dentarg commented Jun 30, 2024

ruby-mqtt 2x slower than paho-mqtt #115

ruby-mqtt 2x slower than paho-mqtt #115

Comments

rgaufman commented May 10, 2018 • edited Loading

njh commented May 11, 2018 • edited Loading

njh commented Mar 9, 2020

dentarg commented Jun 30, 2024

rgaufman commented May 10, 2018 •

edited

Loading

njh commented May 11, 2018 •

edited

Loading