radiusd process vanishes on higher load

[A.L.M.Buxey at lboro.ac.uk]

Hi,

> I'm using a FreeRadius setup (V 1.1.3-3 from Debian etch) with the MySQL 
> Backend for authorize and accounting. RADIUS packets are coming from 
> another company, which /probably/ has a Proxy for their delivery front-end 
> servers.
> This setup usually works fine, but on some days I see spikes in the log, 
> and the radiusd process vanishes; needing a restart.

1.1.3 (even with a few debian patches) is very very old.  I would
advise, if you want to stick with 1.1 train, to upgrade to 1.1.7
to fix several issues and bugs.  2.0 train - 2.0.4 would be a wiser
investment of time.

> I've set:
>   max_request_time = 5
>   cleanup_delay = 5
>   max_requests = 4096
>   sql.conf: num_sql_socks = 256
>   nas: nastype = other

why did you increase the default max_requests from the defaul 1024?

> Usually, there will be a few 'stop packet with zero session length' log 
> lines, about 1 per second or fewer.
>
> At problem/spike time, I see a lot more of these 'stop packet with zero 
> session length' stuff (about 30 per second); and this too:
>
> Info: The maximum number of threads (256) are active, cannot spawn new 
> thread to handle request

this sounds like the number of requests coming in are overwhelming you SQL
server - its not fast enough to keep up with realtime RADIUS accounting
packets. this usually happens suddenly - either when tables get to
certain size or when the number of incoming packets gets to something
reasonable.  so, options to make the SQL faster

1) change the engine - eg use innoDB for the SQL if you use MySQL
2) optimise and add keys etc to tables
3) stop using the DB for realtime accouting - use radsqlrelay/radrelay
for 1.1 or the detail module in 2.0 - the server can then deal with realtime
stuff and the accounting gets moved into 'almost real time' without
affecting the server.

> Is this just a very long spike (can take 10 minutes to a few hours);
> Are there any counters in FR that I could track? (like packets per second, 
> etc.)

if its built with SNMP support you could always use eg RTG/MRTG to graph
the counters...

alan