Yeah, well, Mr. H knows what I think of his *adequate* hardware...  sort of
"please don't shoot the piano player, he's doing the best he can".  But
let's look at things...

Seriously, there's a quantum difference here - both in expectations,
capacity and traffic.

First off, I believe that ntop should be run on a dedicated box for the
dedicated purpose of monitoring traffic.  Both for practical reasons
(contention) and security (if somebody compromises the box through an ntop
problem, all they get is ntop, vs. all the other services).  (Not that I'm
aware of ANY problems with ntop security, but I do understand - and have
documented - it's limitations.

Hardware wise, there's the P3 vs P difference - while ntop shouldn't make
too much use of the new P3 instructions, there are core differences (clocks
per instruction), such that the difference in capability is much more than
the raw 4.4x difference.  (look here:
http://www.tomshardware.com/cpu/20030217/)

As part of the difference, there's also a major difference in memory
bandwidth.  A P3-733 is probably running PC133 memory.  A P-166 probably
runs 66Mhz EDO  See here
http://www.dewassoc.com/performance/memory/how_to_id_pc133.htm for an
interesting article.  The amount of bits per packet is the same, so there's
at least another 2x difference... although caching helps, so let's call it
5x difference in raw processing.

There's also differences that affect overall performance in the Linux kernel
(and I assume FreeBSD) for P vs P3... MTRR for one.

There's also interesting stuff on context switching - dated but interesting
that is - at
http://www.cs.berkeley.edu/~laik/projects/os_benchmarks/publications/usenix1
996.bench.pdf.  It seems to say that the box actively running ntop and thus
doing frequent context changes between a small # of processes will perform
significantly slower under FreeBSD (3.0 albeit) than Linux (1.2!).

There are also gcc options - benefits, etc. unclear - see
http://gcc.gnu.org/onlinedocs/gcc/i386-and-x86-64-Options.html#i386%20and%20
x86-64%20Options.  There are alignment differences P vs P3, which might
impact us because of the unsigned long long counters used.  Somebody might
try -malign-double.


Anyway, put it all together and while you clearly don't explain everything
about the differences, it is a big start.

Other things you need to look at are what that 128K or 10M stream is - the #
of packets is going to be more important than the # of bits per second.


-----Burton



-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] Behalf Of David
Touitou
Sent: Monday, November 10, 2003 7:48 AM
To: [EMAIL PROTECTED]
Subject: Re: [Ntop] openbsd 1st comments: zombie, much cpu


Burton Strauss wrote:

> Given you're both playing with the fxp driver, and the only diff
> seems to be D_P, it's worth a try, right?

And CPU/bandwidth...

I'm sniffing 10Mbps bandwidth with a P3-733.
I consider CPU usage (30%) not that big.

Stanley is sniffing less than 128kbps with a Pentium 166...
He finds 30-50% CPU usage too high 8-/

David.
_______________________________________________
Ntop mailing list
[EMAIL PROTECTED]
http://listgateway.unipi.it/mailman/listinfo/ntop

_______________________________________________
Ntop mailing list
[EMAIL PROTECTED]
http://listgateway.unipi.it/mailman/listinfo/ntop

Reply via email to