[Tile-serving] openstreetmap-carto benchmark results
kakrueger at gmail.com
Thu May 16 22:10:25 UTC 2013
On 05/16/2013 03:27 PM, Paul Norman wrote:
>> From: Kai Krueger [mailto:kakrueger at gmail.com]
>> Cc: tile-serving at openstreetmap.org
>> Subject: Re: [Tile-serving] openstreetmap-carto benchmark results
>> I am not entirely sure I buy your conclusion that once put on yevaud or
>> orm, the difference would be more towards 12% than 22%. You cite numbers
>> of 500 - 700 IOPS that are done in osm.xml. From just those numbers it
>> is hard to say if the 700 IOPs are limited due to CPU or IO bandwidth
>> (after all, it only has 4 cores compared to yevauds 8 - 16 (with
>> hyperthreading) cores, using something like 14 render threads), but
>> yevaud likely has higher IO performance (currently the numbers are ~3k/s
>> average and a max of ~6k/s and it is mostly CPU bound, so probably could
>> do more, even on the single old "slow" 320 SSD). Therefore my guess
>> would be (although it needs closer analysis) that on yevaud or orm, you
>> would see differences closer to your fully cached regime of 22%.
> On the other hand, the smaller sets had an iops of 0 because the data was
> all in cache because it was rendering the same tiles over and over again.
> Yevaud doesn't repeat the exact same list of tiles, making me believe that
> the longer list should be closer to what we'd see on it.
An extremely fast disk system behaves as if everything is in cache.
Where the SSDs lie on the scale between RAM and slow HDDs with respect
to this issue, is at the moment just a guess and will determine if you
will see more of a slow down in line with the larger or smaller set of
files. My guess would be that it lies closer towards the side of all
cached, especially if you have enough rendering daemons, but that is
indeed just a guess and may well be wrong.
But either way, if we do move over to a multi server setup, it should
give the breathing space to switch over now, while at the same time
continuing to figure out how to regain the performance loss with carto.
> For reference, the smaller test a list from about 6 minutes of yevaud's logs
> and the longer list about an hour.
More information about the Tile-serving