package backoff

  1. Overview
  2. Docs
Exponential backoff mechanism for OCaml

Install

Dune Dependency

Authors

Maintainers

Sources

backoff-0.1.1.tbz
sha256=00bea311b227b1bc0a55879da4dce38b1fd8447b4e4e2cf193a6953737ac9f03
sha512=c9c06632c49598831b600eba63a2070fc406c7554a0c8f60807d243b7cd3d996a8bfc9c4d90428c5b1765a2b61ddc26403dc9cdb2b8a6686249f1fdafcec81cb

Description

Published: 28 Sep 2024

README

README.md

API reference

backoff - exponential backoff mechanism

backoff provides an exponential backoff mechanism [1]. It reduces contention by making a domain back off after failing an operation contested by another domain, like acquiring a lock or performing a CAS operation.

About contention

Contention is what happens when multiple CPU cores try to access the same location(s) in parallel. Let's take the example of multiple CPU cores trying to perform a CAS on the same location at the same time. Only one is going to success at each round of retries. By writing on a shared location, it invalidates all other CPUs' caches. So at each round each CPU will have to read the memory location again, leading to quadratic O(n²) bus traffic.

Exponential backoff

Failing to access a shared resource means there is contention: some other CPU cores are trying to access it at the same time. To avoid quadratic bus traffic, the idea exploited by exponential backoff is to make each CPU core wait (spin) a random bit before retrying. This way, they will try to access the resource at a different time: that not only strongly decreases bus traffic but that also gets them a better chance to get the resource, at they probably will compete for it against less other CPU cores. Failing again probably means contention is high, and they need to wait longer. In fact, each consecutive fail of a single CPU core will make it wait twice longer (exponential backoff !).

Obviously, they cannot wait forever: there is an upper limit on the number of times the initial waiting time can be doubled (see Tuning), but intuitively, a good waiting time should be at least around the time the contested operation takes (in our example, the operation is a CAS) and at most a few times that amount.

Tuning

For better performance, backoff can be tuned. Backoff.create function has two optional arguments for that: upper_wait_log and lower_wait_log that defines the logarithmic upper and lower bound on the number of spins executed by {!once}.

Drawbacks

This mechanism has some drawbacks. First, it adds some delays: for example, when a domain releases a contended lock, another domain, that has backed off after failing acquiring it, will still have to finish its back-off loop before retrying. Second, this increases any unfairness: any other thread that arrives at that time or that has failed acquiring the lock for a lesser number of times is more likely to acquire it as it will probably have a shorter waiting time.

Example

To illustrate how to use backoff, here is a small implementation of test and test-and-set spin lock [2].

  type t = bool Atomic.t

  let create () = Atomic.make false

  let rec acquire ?(backoff = Backoff.detault) t =
    if Atomic.get t then begin
      Domain.cpu_relax ();
      acquire ~backoff t
    end
    else if not (Atomic.compare_and_set t false true) then
      acquire ~backoff:(Backoff.once backoff) t

  let release t = Atomic.set t false

This implementation can also be found here, as well as a small benchmark to compare it to the same TAS lock but without backoff. It can be launched with:

dune exec ./bench/test_tas.exe > bench.data

and displayed (on linux) with:

gnuplot -p -e 'plot for [col=2:4] "bench.data" using 1:col with lines title columnheader'

References

[1] Adaptive backoff synchronization techniques, A. Agarwal, M. Cherian (1989)

[2] Dynamic Decentralized Cache Schemes for MIMD Parallel Processors, L.Rudolf, Z.Segall (1984)

Dependencies (2)

  1. ocaml >= "4.12"
  2. dune >= "3.3"

Dev Dependencies (3)

  1. odoc with-doc
  2. domain_shims >= "0.1.0" & with-test
  3. alcotest >= "1.7.0" & with-test

Conflicts

None

OCaml

Innovation. Community. Security.