[OSRM-talk] Using OSRM linked into other code?

Antonio Moratilla Ocaña antonio.moratilla at gmail.com
Fri Nov 7 20:36:29 UTC 2014


Hello

A year ago I was in the same situation, but using Java.

The found the problem was not about OSRM server but connection overhead in
my own program. Creating a new http connection for each request was simply
not an option: i could manage to get about 20 request at max. I tried to
use external todos like wget and curl, but that didn't worked at all.

The solution in my case was to use apache httpclient library in concurrent
mode, with resources for about 12 concurrent conections in 8 concurrent
threads.

With that configuration I managed to pass from as much as 20 request per
second to more than 900 request per second.... At that speed, OSRM was
using about 100% CPU time, so it worked full speed... :)

Maybe this could help you... If you need code let me know

Kind regards

El viernes, 7 de noviembre de 2014, Stephen Woodbridge <
woodbri at swoodbridge.com> escribió:

> I'll answer a bunch of the reply's here:
>
> 1. we do pre-compute a distance matrix and use that already but if you
> have a situation like:
>
> o----------t----------u--------v----->
>            |          |        |
>            B          C        |
>            |          |        |
> A----------x----------y--------z---D---->
>
> and you want the route A-B-C-D if you use a precomputed distance matrix
> you get a path A-x-B-x-y-C-y-z-D (depending on where B and C are in those
> segments (ie: the vehicle makes a u-turn at B and C) when we want a route
> like A-x-B-t-u-C-y-z-D. OSRM will generate the later route if you ask for
> the route A,B,C,D with via points. So a simple distance matrix does not
> work well.
>
> 2. The performance issue is not with the C++, we get basically the same
> performance using Perl (GET) or curl at the command line, or curl from C or
> from c++.
>
> 3. I will look at the node-osrm code. I remember seeing that posted, but
> had forgotten. Thanks for the reminder of that.
>
> 4. I am some what stuck on an older version of the source code because I'm
> not in a position to upgrade my server OS and packages. :( So this is
> somewhat problematic for me at the moment.
>
> Anyway, lots of great ideas. I appreciate them all and will be digging
> into them over the weekend.
>
> Best regards,
>   -Steve
>
> On 11/7/2014 12:46 PM, John Firebaugh wrote:
>
>> Hi Steve,
>>
>> Recent versions of osrm-backend build a library which you can link
>> against. See https://github.com/Project-OSRM/node-osrm/ for an example.
>>
>> cheers,
>> John
>>
>> On Fri, Nov 7, 2014 at 7:13 AM, Stephen Woodbridge
>> <woodbri at swoodbridge.com <mailto:woodbri at swoodbridge.com>> wrote:
>>
>>     Hi,
>>
>>     I seem to remember a while back that there was a discussion about
>>     the possibility to embed the OSRM routing engine at the code level
>>     rather than doing HTTP requests to a server.
>>
>>     I now find myself in a position that this would be desirable to do.
>>     I have a small coverage area like a city, but I'm getting killed by
>>     the overhead of formatting requests as strings, making a socket
>>     connection to osrm-routed, parsing the responses, etc. Making local
>>     requests my server this is taking 4-500 ms per request.
>>
>>     Basically, I'm doing viaroute requests with 2-100 via points. 99% of
>>     the time all I need to know is the travel time.
>>
>>     Since I'm developing in C++, I thought it might be easy and much
>>     faster to instantiate the routing engine and then have a simple
>>     interface where I can pass a container of points and get back the
>>     travel time for that route and/or the path coordinates. But I could
>>     live without the coordinates if I had to.
>>
>>     Has anyone done this already? Can you share?
>>
>>     I have started digging through the source to see if I can do this,
>>     but working my way in from osrm-routed or Tools/simpleclient.cpp the
>>     code is very entangled with all the http request/response stuff that
>>     I would ideally like to avoid. So far the most promising path looks
>>     like using some variant of the simpleclient, but its not obvious if
>>     or how to untangle all the json stuff and simply return a struct or
>>     class to the caller without that. I spent most of yesterday, digging
>>     through this and made a lot of progress just understanding
>>     simpleclient and getting ti to compile and work and get it to actual
>>     return results using a shared memory connection.
>>
>>     A little help in this direction would be appreciated.
>>
>>     Thanks,
>>        -Steve
>>
>>     _________________________________________________
>>     OSRM-talk mailing list
>>     OSRM-talk at openstreetmap.org <mailto:OSRM-talk at openstreetmap.org>
>>     https://lists.openstreetmap.__org/listinfo/osrm-talk
>>     <https://lists.openstreetmap.org/listinfo/osrm-talk>
>>
>>
>>
>>
>> _______________________________________________
>> OSRM-talk mailing list
>> OSRM-talk at openstreetmap.org
>> https://lists.openstreetmap.org/listinfo/osrm-talk
>>
>>
>
> _______________________________________________
> OSRM-talk mailing list
> OSRM-talk at openstreetmap.org
> https://lists.openstreetmap.org/listinfo/osrm-talk
>


-- 

Antonio Moratilla Ocaña - antonio.moratilla at uah.es - Despacho N334

Profesor del Dpto. Ciencias de la Computación - http://www.cc.uah.es
Escuela Politécnica - Informática - http://www.etsii.uah.es
Universidad de Alcalá - http://www.uah.es
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstreetmap.org/pipermail/osrm-talk/attachments/20141107/9e3175ab/attachment.html>


More information about the OSRM-talk mailing list