Hello Andrea and Chris,
Thanks for the quick response.
Quite interesting notes and let me dig deep in to the presentations.
Whats more interesting is the fact that esri and erdas will be
participating in some benchmarking. Will look forward to that.
Surprised that these two are participating, .. tell me thats its TRUE
again.
I'm not against any of these guys, its just the fact that they will
never let you know the performance indicators. If one remembers, the
earlier versions of ArcIMS fails with even 20 simultaneous user if not
properly managed. On the other hand, erdas guys claim themselves to be
the fastest on the world ... well my question is against whom and on
what condition? Does anyone seen any performance numbers published by
erdas? Other day I was reading somewhere that they said it handles
1000's (note 1000's) of concurrent users on a modest hardware... :)
... so I decided to do some test when I had access to one very
recently..
With a very large image requests (1024x800) via WMS on a very large
bluemarble image of 10GB, we sent random bounding boxes through
jmeter.. the results that we got was ... a throughput of 6/sec with
5000 sample request on a modest hardware (note, absolutely NO cache).
Now I dont know where does this 1000's of concurrent users come in? Or
is my understanding, lingo of concurrency is wrong?
On the other hand, GeoServer and MapServer are doing benchmarking for
few years and letting people know the true facts.. hats off.
Getting, back to the original question, the image serving today is
down to image tiles serving for consumer portals and wms (and other
ogc standards) for technical users.. Thats the reason I'm so keen to
see how things are moving...
TMS sounds very promising to me.
Bruce
On Sat, Sep 19, 2009 at 2:56 PM, Andrea Aime <
[hidden email]> wrote:
> Bruce Foster ha scritto:
>>
>> Be it arcserver or erdas, they talk much but not concrete test numbers
>> are revealed for one to make a decision. I saw the last year foss4g
>> presentation on performance, but the numbers dont look correct to me.
>> I was not sure if the test was done purly on random bbox, because the
>> throughput is around
>
> The test were run on my personal notebook, a cheap one too, dual core
> 2Ghz, 4GB memory, 5400rpm hard disk. With a fully cached tile set
> (everything on disk already) I managed to saturate a 100Mbit line
> with both GWC and TileCache (for TileCache to saturate the line I
> had to set it up in mod_python).
> The bboxes were never repeated during the test, it was a set of
> machine generated bounds matching the tile hierarchy and going from
> low to higher zoom (1200 or so different bounding boxes).
>
> Of course if the tiles were not pre-cached the figures would have
> been very different, it stands to reason they would have been
> somewhat slower than the pure WMS sitting on the back of the
> tilecache: get the metatile, slice it up, save to disk, that
> surely takes its toll.
>
>> recently i had access to a erdas server and while doing performance
>> test, it came out shocking low figures. Far less than what geoserver
>> and mapserver as in the foss4g presentation. (well, even though its
>> not apple to apple). quite confusing..
>
> Mind those high numbers where also due to the choice of the data
> set and the rendering option (and to make it possible to run the
> tests a number of times without that ending up requiring hours).
> The big road layer was rendered very close to native resolution and
> without antialiasing, the state layer is small, the tiff was small
> as well and, well, the MrSid one was slow, but MrSid is slow to open
> with almost anything I guess.
>
> The dataset and the JMeter scripts are available for you to try
> them out if you want.
>
>> assuming on a entry level server (intel xenon duel core, 2 gb ram,
>> 500gb 7200rpm disk), one get a throughput of 12/sec with WMS, what
>> will the performance be with TMS?, will that be 10 fold or 50 fold?
>>
>> On a curiosity note, how does google serve? I know they have massive
>> brute force with respect to server power and bandwidth.. but does it
>> serve TMS like tiles?
>>
>> the other subject of performance, the known bottle necks are;
>>
>> 1. Disk read speed, 7200rpm disk vs 15000rpm disk vs SSD disk ...
>> 2. Bandwidth 100 Mbps (100 mega bit per second or 12.6 mega byte per
>> second), this can carry only a certain amount of traffic only.
>> 3. Network card and the routers, switches to which they are attached
>> 4. CPU of the server, xenon was amd etc...
>
> the CPU will bottleneck you with the time it takes to setup the
> cached tile set, but once everything is cached it's pretty much irrelevant.
> When I saturated the 100Mbit local line the notebook
> CPU utilization was not going over 20%
>
> Cheers
> Andrea
>
> --
> Andrea Aime
> OpenGeo -
http://opengeo.org> Expert service straight from the developers.
>
--
Thanks
Bruce
_______________________________________________
Tilecache mailing list
[hidden email]
http://openlayers.org/mailman/listinfo/tilecache
