(RADIATOR) a summary of some recent benchmarking results

[Tariq Rashid]

the following bechmarking  results summary may interest some of you....

keep in mind that these are specific to our own configuration and your
results may differ (eg our auth is more intensive than our acct, simply
because we do much more in our post-search-hook files...)

t

-------------------------------------------



some recent results:


multiple-instance with a proxy:
~~~~~~~~~~~~~~~~~~~~~~

	* proxying adds latency - which signifcantly reduces radiators
ability to handle
		high request rates. this appears to be a weakness of
radiator.
		on my box, the normal critical rate of 17/sec is reduced to
13/sec

	* adding multiple instances does not affect this critical rate
		also adding multiple instances does not change the
		rate of performance degredation

	* a singler server performs better! (????)

	* this apparent weakness has raised the interest of some of the
mailing list users



auth, acct and no-hook:
~~~~~~~~~~~~~~~~~~~~~~~

	* accounting is much quicker than auth (yes, i did confrm that the
		SQL insert and delete are sent to the load balanced DB)

	* the shape of the responce-time/request rate has the same form as
the
		auth request curves - ie degredation in performance is
similar

	* the critical rate for acct packets is about 30 requests per second
		compared with the 15 to 17 per second for auth

	* mixing the acct/auth in equal numbers gives a similar degredation
again,
		with the critical point at about 22 requests per second

We suspected that the hookk-file is soaking up the time ro respond. So I
compared it with a configuration using no post search hook file:

	* no-hook does not affect purely acct requests (acct doesn't use
hook files)

	* auth is quicker - and the critical rate increases to about
22/second. not a great
		improvement. the degredation is the same shape.

	* this suggests that even if we really stream line the
post-search-hook.pl
		we cannot expect a better critical point than 22/second for
		auth requests. - if it takes a month to really sort out the
hook file
		the developmeny cost may be better spent on an additional
		single instance radius server with a hardware load balancer.

	* the base "happy", unstressed response rates are:
		- 0.04s for ACCT (hook, no-hook)
		- just under 0.06s for auth, no-hook
		- about the normal 0.07s for auth, with a hook file

i can talk you through the graphs next week if you wish.

next week i'll start on freeradius - so see the affect of "C" code that is
also multi-threaded. I will also try to find a faster server to run the
radiator on to see if a faster machine affects the above and previous
characteristics.

tariq

--
Archive at http://www.open.com.au/archives/radiator/
Announcements on radiator-announce at open.com.au
To unsubscribe, email 'majordomo at open.com.au' with
'unsubscribe radiator' in the body of the message.