(RADIATOR) explanation of hyperthreading on freebsd performance (multi-instance radiator)

[Dave Kitabjian]

Tariq,

I don't have any data to share but I wanted to raise a question about
the perceived benefit of the "halt on idle" and "100% throughput".

My question is, what is defined as "100% throughput"? Is "top"s
definition of "100%" dependent on that sysctl "halt on idle" setting?
Does "top" consider processing "idle" instructions work?

Another way of attacking the question is, are your figures based on
"top"s CPU utilization figures or based on an empirical count of
Authentications per second? The latter would be more meaningful, I
think.

Dave


> -----Original Message-----
> From: Tariq Rashid [mailto:tariq.rashid at uk.easynet.net]
> Sent: Tuesday, November 23, 2004 9:07 AM
> To: radiator at open.com.au
> Subject: (RADIATOR) explanation of hyperthreading on freebsd
performance
> (multi-instance radiator)
> 
> 
> i know radiator is not a threaded server. however we can run multiple
> instances on the same hardware.
> 
> recent hardwarwe is capabale of what is known as hyperthreading, which
> gives
> the illusion of 2 cpus from one. we can even have 2 physical cpus to
give
> the illusion of 4 logical cpus. the rough figures are that you get
about
> 150% work from a hyperthreaded system which is being taken advantage
of
> fully (not always in opractise).
> 
> this is a topic not well documented so i'll make a go of it and
perhaps
> people from the list can correct me or confirm my explanation.
> 
> however - my own experiements on freebsd 4.7+ (4.8, 4.10) have shown
that
> (radius request handling) performance actually degrades with an SMP
kernel
> on dual xeon's with hyperthreading (HTT) enabled. there is a sysctl in
> freebsd which supresses the migration of processes to a second logical
> cpu,
> keeping the processes on the first logical cpu. this was enabled by
> default
> by the bsd team, i think. the results of "top" confirm this.
> 
> what appears to be happening is that when there are 2 logical cpus and
> only
> one is allowed to take processes, the effective throughput is 60-70%
of
> that
> of a normal single non-htt cpu. hence the performance drop. why is
this?
> the
> 2nd "logical" cpu is busy executing "idle" instructions. it is not
halted.
> 
> recently i have looked again at the matter and found that it is
beneficial
> to allow the second logical cpu to "halt" when idling. this return the
1st
> logical cpu to 100% throughput. effectively. this is the same as
disabling
> hyperthreading. this "halt on idle" is enabled by another sysctl.
> 
> i have not repeated my experiments with this "halt on idle" systl, but
it
> is
> reported to improve performance for both cases - when single threads
are
> run, they are run at 100% and when the worload is busy, both logical
cpus
> are used and the effective throughput is up to about 130%.
> 
>  Does anyone have any experience of this? any insights or
clarifications?
> 
>  Does freebsd 5.3 handle HTT scheduling better (affinity, prevents
> resource
> sharing clashes)? Linux 2.6?
> 
> 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.


--
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.