[Mono-dev] Hacking mono to make it more deterministic

[Rodrigo Kumpera]

Forcing sys.object::hash_code to be zero will result in insertion order to
Dictionary
and cause all operations to be linear.


On Mon, Aug 24, 2009 at 10:13 AM, Lucas Meijer <lucas at lucasmeijer.com>wrote:

> Hey,
>
> At Unity we have this internal testing system where we
>
> 1) play a game, and record all inputs
> 2) can replay that game, based on the recorded inputs
> 3) whenever we do a checkin on our sourcecode, we run all games in the
> testing setup against this new unity build, by making screenshots every
> 5 seconds, and comparing those to the previous run.
>
> This has proven to be very useful, as it finds regressions in our users
> games that we had not anticipated when committing a certain change to
> the unity codebase.
>
> Currently, not all games are playing back deterministically on a single
> unity build. We don't think we'll ever reach full deterministicality (is
> that a word?) for 100% of our games, but the higher we get the number,
> the better we like it.
>
> In order to do this, we mock out things like System.Random,
> DateTime.Now, to return the recorded values.
>
> A frequent case where games fail to run identically is when our users
> iterate over a dictionary which has a custom type as its key. Since the
> dictionary uses GetHashCode() internally, and Object.GetHashCode() is
> based on the memory address of the object, it means that the order in
> which the KeyValuePair's get returned can differ per run.
>
> Naturally our users shouldn't be doing this. But they do, as they don't
> realize that consistent ordering of iterating over a dictionary isn't
> guaranteed. it just works most of the time so most people don't notice.
> I didn't realize this myself untill I started investigating this issue :)
>
> I'm looking for advice on what would be the best solution to this.
>
> So far the best idea I've came up with is to make mono_object_hash just
> return 0. That probably has some severe performance implications, but we
> can live with those. I'm hoping this will only make things go a lot
> slower, but wanted to ask this list:
>
> - if I should expect other problems than slowdown
> - if you can think of a better/different way to make our users code
> deterministic
>
> Thanks, Lucas
> _______________________________________________
> Mono-devel-list mailing list
> Mono-devel-list at lists.ximian.com
> http://lists.ximian.com/mailman/listinfo/mono-devel-list
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.ximian.com/pipermail/mono-devel-list/attachments/20090824/e811ccf4/attachment.html