See [BMS]'s inline -- also consider punctuation and carriage returns in the
future - run-on sentences are difficult to read, esp. for non-native English
speakers.

-----Burton 

-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of
[EMAIL PROTECTED]
Sent: Wednesday, September 07, 2005 12:38 PM
To: [email protected]
Subject: [Ntop] My feedbacks on Ntop

Hello list,

This is my feedbacks on Ntop.

1) - It would be nice if we can have an option into configure to enable or
disable XMLDump, by defaut if libxml2, dome, and very OLD glib1 are not
available, then the XML feature is not functional, that's correct but I
don't understand why "libxmldumpPlugin" and "xmldumpPlugin.so" are still
compilled and installed!! Therefore having an option to enable or disable
this feature could be a good addition to the source, also why you still use
glib version 1.x?

[BMS] Live with it.  The gnu tools don't like missing .o files (they insist
on recompiling each time you 'make' then you get problems from files owned
by root after the make install).  So it's easier to have an empty 'program'.

[BMS] There's no ntop dependency on glib1.  We happily run with whatever is
installed on the box.  Here's my Fedora Core 3 box:

$ rpm -q --whatrequires glib
gdome2-0.8.0-1
pam-0.77-40
gtk+-1.2.10-29.1.1
GConf-1.0.9-13.1
gnome-vfs-1.0.5-18
glib-devel-1.2.10-12.1.1




2) - Running Ntop for home user or small company using Hub work fine because
they have not lot of trafics on their network and Ntop work nice on this
environement but when we talk about medium or large entreprises using
switches (Cisco, FoundryNetwork), then Ntop become a problem. For Ntop to
work, we have to use additional switch features like Netflow for Ntop to see
the trafic or SPAN if we don't have US 3000$ to buy a Netflow card. 

[BMS] Do the math.  NO general purpose computer can keep up with backbone
frame rates.  There's simply not enough memory bandwidth.  10,000 9K packets
per second - where each packet (at best) has to xfer from the NIC to kernel
buffer, and then kernel buffer to user space - is 1.4 Gbps.  Toss in
real-world, where there's at least one extra copy and... Remember: PC3200
maxes out at 3.2 Gbps.  Real-world throughput is less.  And that's without
any ntop processing.

[BMS] Check the stats on the about page.  On a 1 GHz P3, ntop can handle
about 1-2K packets / second (regular, not jumbo).

[BMS] So technologies like netFlow (1) do massive data reduction - 15 9K
packets -> 1 1400 byte flow packet and (2) reduce the per-packet processing
times, because we don't see the internals and so don't do deep packet
inspection.

[BMS] If you are running in the enterprise space, you have to expect to
spend $s for enterprise hardware.  Or do you think a hardware sniffer is
free?


In both way activating Netflow or SPAN will generate lot of trafics on the
server (also on the switch for SPAN) where Ntop is running and even with an
Intel Pro 1000 Ethernet Adapters supporting Jumbo Frame, NAPI, etc, the
server and Ntop are not able to handle the load. Ntop consume 80% of all CPU
resources and all 3GB of RAM + 1GB of swap in a minute.

[BMS] Re traffic, yes.  Any monitoring technology will have overhead.  

[BMS] With Span it's the 100 or 1000 Mbps of the link itself.  With netFlow
it's the flow packets.  If you are so tight on bandwidth that you can't
handle this, then set up a private link (a/k/a a wire) to keep the traffic
off the backbone.

[BMS] Re memory usage, we've cut it way down, but the hard facts are that
there is some overhead for EVERY host we monitor.  Source and destination.
Or you can use filters and switches to cut it down to what you need to
monitor.  Like everything in life, you have to live within your means.

[BMS] FWIW, under Linux, ntop shouldn't use that much CPU.  Under FreeBSD
it's an artificial usage because of the userland threads as explained in the
FAQ.

Since the kernel still need to manage other process, it will automaticaly
kill the process causing the problem by consuming all resources (yes Ntop).

[BMS] ntop should - especially in a production environment - run on a
standalone box.

So which server should I have to run it? A 4 Xeon processors hypertheading
with 50GB of RAM on a RAID SCSI system? Ok there is a PF_RING solution that
ONLY work with old libpcap version and old Linux Kernel with old GCC version
too, so when you use a new Linux system to be update and bug fix, you cannot
use this solution and even applying this solution with old software version
doesn't fix the above problem.

[BMS] Above 1 core per NIC (or maybe 2 + #NICs), ntop won't see any
benefits.  The packet processing threads could max out a core per NIC, and
the web server one, with the rest lightly loading another core.  But ntop
runs OK even on a single processor with modest loads.
  
[BMS] So take it up with the network device driver authors - why are their
drivers so slow?  All we can do is to use what's available to us.  PF_RING
was a research project to see if significant improvements COULD be made.  If
it's so important to you to keep it current, and you are in an enterprise
environment, then consider funding the development effort.

Yes ntop is a great software and I like it, but I'm affraid to explain all
of the above in the entreprise when I've to deploy it.

[BMS] So don't.  You have choices - spend $50K + hardware costs and license
a commercial product or spend less for the admittedly lower functionality of
ntop.  Or, take some of that money and fund development projects to improve
ntop to meet your needs, which benefit everyone...   The benefit of ntop is
not that it's the only solution for everyone, it's that you have the choice!



Gerhard,

_______________________________________________
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