The practical benchmark of classic Laravel queueing mechanisms: redis, beanstalkd, database (MySQL, postgre)

Intro

Laravel provides several different mechanism for job queueing. I will cover: Redis, Beanstalkd, Database (MySQL, PostgreSQL)

For the benchmarking, I created a new Laravel 5.5 project. The benchmarks were performed on 2 VPS:

Amazon EC2 c4.large

  • Intel(R) Xeon(R) CPU E5-2666 v3 @ 2.90GHz
  • 2x VCPU, 4 GB RAM
  • EBS, SSD storage (with burst credit > 0)

Forpsi smart VPS, low budget server

  • Intel(R) Xeon(R) CPU E5-2650L v4 @ 1.70GHz
  • 1x VCPU, 1 GB RAM
  • SSD storage, without IO throttling

Both servers are running the same software, latest Ubuntu 16.04.3 LTS. MySQL version: 5.7.20, PHP 7.1.12-3 with opcache, PostgreSQL 9.5.10, Redis 3.0.6, Beanstalkd 1.10.

The packages are in the default configuration.

Method

  • There are always 10 workers (processing the jobs) started before the test.
  • Jobs being processed are empty - doing nothing. This way I measure the job scheduling system overhead as the workers are competing for the jobs.
  • 10 000 Jobs are pre-generated on the job queue. During the pre-generation the Laravel is in the maintenance mode - not processing the jobs.
  • After the jobs are generated Laravel is switched to normal mode - jobs processing will start.
  • Sample the size of the job queue every 0.5 seconds. Measure the time needed to process 10 000 jobs.
  • The whole test is run 10 times.

The overall test produces average jobs per second.

Note the Database queue driver has been slightly modified to overcome the current limitation of running one job multiple times. The original implementation fails to delete some of the finished jobs from the database under high load. More on this problem in the next blog post.

Results

The box plot below is a box chart visualizing the statistical distribution of the jobs per second in 10 runs for each configuration.

Benchmark results

Forpsi:

Method Jobs per second
Beanstalkd 1109.83
Redis 872.47
DB-mysql 190.47
DB-pgsql 206.39

Amazon C4 offers 2 thread cores and the processor is a bit faster than forpsi:

C4.large

Method Jobs per second
Beanstalkd 2600.96
Redis 1905.53
DB-mysql 452.04
DB-pgsql 273.13


Burst credit

When testing on Amazon EC2 it’s important to realize the EBS storage drives are subject to IO throttling.

Each drive has so-called burst credit. Intensive IO operations decrease the credit over the time. Conversely, during low IO usage, the credit restores slowly.

When credit reaches 0 the IO throughput decreases significantly (10% ish). I performed the benchmarks only with non-null burst credits to measure the best possible performance.