Hierarchical Timing Wheels for Clojure and ClojureScript
Switch branches/tags
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.


Ri Gui (日晷, Sundial)

Build Status Clojars GitHub license

Hierarchical Timing Wheels for Clojure and ClojureScript. This is a general purpose timer implementation that scales. Its performance can be tuned via parameters.

Timing wheels is designed for scenario of large scale task scheduling. According to current benchmark it provides way better accuracy for large amount of tasks emitted in short range.


Base Usage

Start a timer:

(require '[rigui.core :refer [start later! every! stop cancel!]])

;; starting a timer with arguments:
;; tick size: 1ms
;; bucket per wheel: 8
;; handler function: some-handler
;; thread pool for running tasks: some-executor (for JVM only)
(def timer (start 1 8 some-handler some-executor))

Schedule some task/value for later:

;; schedule some value :a with delay 1000ms
;; timer's handler function will be called with :a in 1000ms

(def task (later! timer :a 1000))

Note that values to scheduled within a tick will be delivered to handler function at once, and executed on current thread.

The handler function will be called with task value as arguments when the task expires. The run a custom task, you can start the timer as:

(def timer (start 1 8 (fn [f] (f)) some-executor))

(later! timer #(println :a) 1000)

The schedule function returns a promise-like object, which maintains the execution result of handler function and task value. deref, blocking deref and realized? are supported.

Cancel a task:

(cancel! timer task)

Schedule some task/value for a fixed interval. The value will be delivered to handler function in fixed interval.

(every! timer :a
  1000 ;; initial delay
  500 ;; interval

Stop the timer:

;; calling stop on timer returns all the pending tasks
(stop timer)

Once a timer is stopped, it no longer accepts new task.

Delayed Channel

This library also provides an experimental core.async channel that pop the value after some delay.

(require '[rigui.async :refer [delayed-chan delayed-value]])

(def c (delayed-chan))

;; the value will be available in 1000 milliseconds
(>!! c (delayed-value :a 1000))

;; blocked until 1000 milliseconds later
(<!! c)

Performance comparison

In the bench/ source directory, I made a bench script to compare performance of Rigui and ScheduleThreadPoolExecutor. With the (thoughput 50000), I try to schedule 50000 tasks that will be emitted within 5 seconds. This function will prints its result to stdout:

Testing enqueue time for JVM timer and Rigui
Enqueue 50000 tasks into JVM timer
"Elapsed time: 89.618297 msecs"

Enqueue 50000 tasks into Rigui timer
"Elapsed time: 591.834465 msecs"

JVM ScheduledThreadPoolExecutor error
avg:  88.53232
max:  181
p99:  181
p95:  175
p75:  136
p50:  90

Rigui error
avg:  18.88236
max:  234
p99:  185
p95:  116
p75:  14
p50:  3

The results indicates:

  • Rigui is slower at enqueuing tasks, which might because rigui uses STM internally
  • Rigui has better accuracy when tasks emitted in a short time. Rigui has way smaller error for 75%-95% tasks.

Under the hood

This library is created with inspiration from this blog post about Kafka's timer improvement. And you can find more material about various timer implementation from this paper.

Generally speaking, the timing wheels implementation trades accuracy for throughout. It aggregates timer tasks into a few buckets and runs these buckets with a small number of actual timers. So timing wheels is specifically optimized for scenarios with a large number of timer tasks, also to be triggered in a narrow window. For instance, tracking request timeout in asynchronous environment, or ping/pong timeout for a large number of connections.

For a single task, the hierarchical timing wheels might create multiple timers (depends on task delay, tick size and bucket count). So there is minor overhead when dealing with small number of tasks.

The accuracy is controlled via tick parameter. Tasks to be triggered within [ tick * n, tick * (n+1) ) will be put into the same bucket and be triggered at the same time theoretically. If you want better accuracy, you may set tick to a small value such as 1ms. But note that finer tick may lead to more actual timers and reduce the throughout.

The second parameter bucket-count decides how many buckets will be on a single wheel. For this hierarchical wheels, it also decides the tick side on the nth wheel (where n > 1, n counts from 1), that is tick * (bucket-count**(n-1)). The larger bucket-count you set, the more internal timers you will have. But in most case it's still less than your task count significantly.

Let tick = 1, bucket-count = 8, the wheels could be visualized like:


A task with delay 5 will be put onto the first wheel, while a delay of 350 will be put onto the third.

If you know good free software to draw this please kindly let me know.


Copyright © 2015-2016 Ning Sun

Distributed under the Eclipse Public License either version 1.0 or (at your option) any later version.