[Tile-serving] openstreetmap-carto benchmark results

Andy Allan gravitystorm at gmail.com
Fri May 17 08:53:35 UTC 2013

On 17 May 2013 05:36, Paul Norman <penorman at mac.com> wrote:
> On the other hand, the switch to carto has never been about
> speed, it's about having a sane way to edit the stylesheet.

Oh indeed, and we mustn't lose sight of this. The performance testing
is largely to ensure that things won't blow up on day one when we put
it on the servers. I think we've shown that it won't blow up, but
yevaud might still struggle. I think the rule of thumb is that we
don't need to think twice if it's within 10%.

> I'm working at it from a different angle - I'm trying to identify
> performance improvements, most of which would likely apply to osm.xml as
> well.

I'm interested to know two things

a) What layers are *slower* than the XML version. I suspect the
amenity-points and amenity-symbols ones are slower due to the
attachments issue, but I want to avoid a few hours of refactoring if
it only nets a 0.1% improvement. So identification of slower layers is
very useful.

b) Which layers are just *slow* in both styles. Paul's graphs show a
tremendous jump at zoom 13 which I find really interesting, since it's
not obvious what features being drawn at that zoom level could account
for such a large increase. It might be a combination of things, but
obviously if it's one layer in particular, that would be worth
focussing on, rather than optimising layers at random.

In the meantime, I continue to work on openstreetmap-carto with an eye
to making it easier to work with as my first priority.


More information about the Tile-serving mailing list