HTTP Persistent Connection in osrm-routed

[tmpsa]

We are running osrm-routed on a local box, and send thousands of requests to it. Then our client runs out of 'Ephermal Ports', probably because a new HTTP/TCP connection is created for each request.

Does osrm-routed check and use the HTTP header 'Connection: keep-alive' (which should control the SO_KEEPALIVE socket option)? Or does it close the connection immediately? Our client (using Java SDK) sends the header in the request, but the response does not appear to have it.

How, exactly, does osrm-routed handle Persistent Connections?

[drieks]

Hi,

we have the same setup here. you can work around this on linux by setting this:

# fin timeout, default is 60s, close an TIME_WAIT socket after 2 seconds
echo 2 > /proc/sys/net/ipv4/tcp_fin_timeout

# recycle TIME_WAIT sockets faster:
echo 1 > /proc/sys/net/ipv4/tcp_tw_recycle

[flohoff]

Still a HTTP/1.1 keep alive session would be much preferred. In the end OSRM is that fast that the whole process is latency bound and so tcp handshakes etc add up.

I am having the same issues and tweaking the tcp stack to work around missing features is at most a temporary solution :)

[emiltin]

depending on your use case - but for maximum speed you can also use osrm as a lib and compile it into your own app

[DennisOSRM]

or use the NodeJS bindings

[tmpsa]

emiltin: We would love to have osrm as a lib, but last time we checked (in 2012) it was not portable to Windows; it's heavily dependent on a Unix environment. We need to run it on both platforms. So, a portable osrm would be great!
(And making software portable is an effective way to find some bugs, too).

[DennisOSRM]

We do have Windows support for a couple of months now. The latest binary from develop branch is available here: http://build.project-osrm.org

[flohoff]

I am running mostly perl pulling geometrys from a postgis - running osrm and putting back the geometry into the postgis. OSRM working as library with perl bindings would be an option.
A unix domain socket or UDP interface to talk HTTP to the OSRM, or persistent connections would be a preferable way IMHO. Otherwise you would need to write OSRM lib bindings for every little language out there.

[emiltin]

@woodbri created postgresql extentions/tools for interfacing with osrm: https://github.com/woodbri/osrm-tools#postgresql

[mehany]

Hi, does this issue still persist? @drieks do I still need to apply the work around you suggested? and where?

[drieks]

Hi @mehany, sorry but I can't tell you because I'm not working on this for some time...

[TheMarex]

The example HTTP server is really not suited very well for production. We mostly keep it around for debugging and make small deployments easier. I would highly recommend to use the node bindings or the C++ API if you really have production requirements.