asterisk users - Jan 2004 - P4 processor with Hyperthreading and Asterisk

Hello,

I read an interesting article on 2cpu.com that talks about the performance
differences between a P4 processor with Hyperthreading(HT) enabled and
disabled.
http://www.2cpu.com/articles/ht_explored/index.html

The bottom line of the article is that processor intensive applications
greatly benefit from HT being enabled, but memory intensive applications can
actually see a slight decrease in performance if HT is enabled. I have tried
Asterisk on the same machine with both HT on and off and haven't noticed
much difference in overall performance. 

Is Asterisk's high memory usage canceling out most of the performance gained
by using HT? 

And the best question is should I spend my money on a dual Athlon machine
instead of a P4C with HT?

Thanks,

MATT---

On Wed, 2004-01-07 at 07:24, mattf wrote:
> Hello,
> 
> I read an interesting article on 2cpu.com that talks about the performance
> differences between a P4 processor with Hyperthreading(HT) enabled and
> disabled.
> http://www.2cpu.com/articles/ht_explored/index.html
> 
> The bottom line of the article is that processor intensive applications
> greatly benefit from HT being enabled, but memory intensive applications
can
> actually see a slight decrease in performance if HT is enabled. I have
tried
> Asterisk on the same machine with both HT on and off and haven't
noticed
> much difference in overall performance. 
> 
> Is Asterisk's high memory usage canceling out most of the performance
gained
> by using HT? 
> 
> And the best question is should I spend my money on a dual Athlon machine
> instead of a P4C with HT?
I was reading an article that showed up on the inquirer the other day
comparing the HT with AMD's new dual core approach. The basic comments
they talked about was that the HT is only useful for cache misses and
therefore your dataset for at least one of the threads needs to sit in
cache. If you have a double cache miss, then you stall both sides of the
HT and therefore have even higher contention for the memory bus. AMD's
dual core chips will have their own cache per CPU and will therefore
have a less likely problem of having both cores having a cache miss
unless AMD chooses to make the cache size small. 

The rest of the article basically mentioned that on a set of apps that
used a fair amount of memory, more CPUs would probably be better.
Neither Intel's HT, Intel's dual CPU, or AMD's dual core would be
able
to beat a dual AMD with two sets of memory as is suggested in the
hypertransport layout. This is due to the fact that each CPU has it's
own cache and it's own memory controller for the memory bank sitting
next to it. There would only be contention for the memory if it had to
cross the hypertransport and ask the other CPU for the memory space.

Now that you have a bit more information, choose away at what you want
to do.
-- 
Steven Critchfield  <critch@basesys.com>

mattf wrote:
>Is Asterisk's high memory usage canceling out most of the performance
gained
>by using HT? 
>
>  
>
In my experience, I've found that hyperthreading is a major problem on 
Linux Kernel 2.4.   If I turn on hyperthreading and start to load up an 
Asterisk box we get ratty sounding audio along with pops, clicks and 
ticks, but if I disable hyperthreading (noht on the kernel line in 
lilo/grub) and all of those anomalies go away and the box scales a whole 
lot higher.

The problem is that the 2.4 Kernel sees the one hyperthreaded processor 
as two totally separate processors, so the 2.4 kernel tends to mess up, 
badly, when scheduling the processing.

This problem was delt with in the 2.5 kernel and i've been told by quite 
a few trusted people that it works flawlessly, but i'm not about to run 
a unstable kernel (or even anything less than like 2.6.14) on any of my 
production boxes.


Jeremy McNamara