Re: [PATCH/RFC] gdbserver: Add command line option to not use SO_REUSEADDR

[Alan Hayward <Alan dot Hayward at arm dot com>]

> On 14 Feb 2019, at 14:33, Tom Tromey <tom@tromey.com> wrote:
> 
> Alan> When running with -j55 on HEAD
> 
> Tom> Jealous.
> 

:) But, the disadvantage of large server boxes is building and running
everything remotely, requiring network access.  You gain something, you
loose something.


> Also, I was wondering if this speeds things up.  When I last looked at
> this (a long time ago now), parallelizing the test suite didn't help
> past -j12 or -j15 or so, because there were a couple of very long tests
> that were "tent poles" after this.  As machines get more CPUs, though,
> perhaps we should be looking into splitting those up.
> 

Gave this a quick try on make check for gdb.

-j55:  1m42.106s
-j40:  1m42.595s
-j30:  1m44.216s
-j20:  1m56.600s
-j10:  3m16.514s
-j5:   6m0.417s

There’s clearly a tailing off going on above 20. Although, under two minutes
seems reasonable enough to me. I couldn't think an obvious way to quickly
see what the longest running tests were. Could also be test suite overhead.
Suspect results might be more extreme with the gdbserver boards.


The socket clash is still going to happen too at -j10, will just be less
frequent.


Alan.