Re: [PATCH] travis-ci: run previously failed tests first, then slowest to fastest

From: Jeff King <peff@peff.net>
Date: Jan 20, 2016, 01:53 UTC
Message-ID: <20160120015310.GB24541@sigill.intra.peff.net>
In-Reply-To: <20160120002606.GA9359@glandium.org>

On Wed, Jan 20, 2016 at 09:26:06AM +0900, Mike Hommey wrote:

Show 6 quoted lines

> Having recently looked into this, the relevant travis-ci documentation
> is:
> https://docs.travis-ci.com/user/ci-environment/
> 
> which says all environments have 2 cores, so you won't get much from
> anything higher than -j3.

FWIW, I settled on "-j16" on my 8-core (well, hyperthreaded quad-core) machine after experimenting. That's running the tests on a RAM-disk, though. On a slower filesystem where fsync() actually does something, you're going to get a lot more I/O stalls, and want a bigger CPU to process multiplier.

Here are actual numbers from my machine:

  -j | time (user+sys)
  ---+------------------
   1 | 5m18s (41s+17s)
   2 | 2m24s (41s+14s)
   4 | 1m15s (46s+13s)
   8 | 0m56s (65s+18s)
  16 | 0m53s (76s+24s)
  32 | 0m57s (78s+25s)

Note that the CPU-second times will go up with more threads because of the frequency scaling.

So yeah, -j3 might not be that unreasonable, depending on the filesystem response times.

> The https://docs.travis-ci.com/user/ci-environment/ document says the
> default is "sudo: false" for repositories enabled in 2015 or later, which
> I assume is the case for the git repository. "sudo: required" is the
> default for repositories enabled before 2015.

Thanks. The document I saw used the word "recognized", and I didn't quite know what they meant. We just enabled this a month or two ago, so we should be running on the new format.

-Peff