Let's have another vendor weigh in :) See my comments in line.
-----Original Message-----
From: THolman@toplayer.com [mailto:THolman@toplayer.com]
Sent: Friday, June 03, 2005 8:25 AM
1) Gigabit performance is irrelevant; it's the packets per
second that count. Vendors cheat and claim 1Gb performance
based on large packet sizes (not real world), or just add up
the sizes of all their interfaces.
It would be nice if there was a standardized IPS performance test with
regards to packet size, traffic mix, etc. I don't see that happening
unless
ICSA does it for the NIPS certification. This would cut down on the shady
performance numbers that Tim refers to.
2) In PC architecture, the PCI bus is the bottleneck, not
the processor.
That depends on what you are doing with the processor. If you are doing
pattern matching in the CPU you could run out of CPU well before you run
out
of bus capacity. A PCI bus has a theoretical limit of 1.05 Gbps. A 16 lane
PCI-Express bus is 80 Gbps. Several vendors are already shipping 10 Gig
PCI-Express cards.
3) An Intel processor has a large instruction set designed
for workstation/server performance and GUI operations, and
not for packet processing.
I would say that the processor designers didn't have any specific tasks in
mind. It is a general purpose processor.
4) An ASIC has a tiny instruction set in comparison,
designed for a specific task. So, a 3.2Ghz Intel processor
forwarding/processing network traffic is on a par with a
133Mhz ASIC designed to do the same thing.
I'm not an ASIC guy so I will take your word for it on the comparison :)
5) Processors can only do one thing at once. Thus, a
networking device with several processors installed in
parallel (ASICs OR Intel) is far more effective than a box
with a single/dual processor.
More processors gives you more flexibility in what gets processed where.
6) Hard disks do not slow down performance. They lower
reliability as fail all the time (!). RAID would help, but I
don't think any security vendor offers a RAID array as an
integral part of their appliance, so cut to the chase, get
the HDD off the inline unit and place on a separate
management machine so we have a reliable distributed
architecture that isn't put at risk by HDD failure. On the
same note, dual fans and power supplies also need to be considered.
Hard drives do fail, no question there. I definitely disagree with your
statement about vendors not having RAID. There are definitely vendors
(other
than us) who have drives in RAID configuration, both 1 and 5. I am not
sure
taking the drive off the device makes for a more reliable distributed
architecture. What if the link from the IPS to the Management machine goes
down or the Syslog server dies? What if the hard drive in the Management
machine fails? :) With no drive on the IPS your space to store events,
system data, etc, is somewhat limited. How long before you have to start
overwriting event data on the IPS?
Same goes for dual fans and power supplies. There are vendors (again other
than us) who have dual fans and hot swappable power supplies. Although
these
are generally found in the 500 mbps and up ranges.
Don't forget fail open NIC's and bypass devices. Most vendors (including
ASIC IPS') have them, at least as an option. If not having a hard drive is
the path to reliability then why do vendors without hard drives have fail
open NIC's? Because other components can and do fail as well.
7) Single-processor machines can easily FORWARD 64-byte
packets at 'multi-Gig' speeds. They can do this as the
processor doesn't have to do anything with them. As soon as
you add intensive operations to the packets in question,
bearing in mind there is only a single CPU that can only do
one thing at once, you introduce LATENCY plus reduce pps
performance DRASTICALLY. This is where a parallel processing
architecture comes into it's own and takes leaps forward over
what a single-CPU box can do.
You are assuming that the CPU is doing the packet processing. Many vendors
are using network content accelerators and other processing cards to
offload
the CPU intensive operations.
In conclusion:
A box with one or two ASICs in is easily outperformed by a PC
with the latest Intel processor, fast network cards and a
good chunk of memory.
However, the PC is more prone to hard disk failure, which is
why you should never put one inline if uptime is critical.
A box with several ASICs in will outperform ANY PC-based
solution, and ANY ASIC solution that relies only on one or
two processors.
But at what cost in terms of price per Gigabit and flexibility? Adding new
functionality to software is pretty easy....
..and one comment to Ed with respect to McAfee/TippingPoint
>both products really don't care if you have every signature and then
>some on.
Yes they do. If you turn on every signature check with these
IPS's, pps performance slows to a mediocre dribble...
They do care. Look at some of the product reviews and you will see that
vendor X has 2000 rules / filters / signatures but only 500 are on by
default. I've witnessed a couple of ASIC IPS' that were brought to their
knees when asked to store the offending packets. What about storing the
TCP
stream involved with an event? Customers are asking about this...
Inline devices should NOT rely on REGEX signatures - by
nature, string searching is very resource intensive and best
left to a nice fast offline IDS running on an up-to-date PC
platform, where latency is not going to be an issue...
There are PC platform IPS on the market that are under 100 microseconds
that
do pattern matching.
Hope this helps - this isn't an all out war ASIC-based vs
PC-based, it's a question of architecture and suitability for
the job in hand!
Definitely an interesting thread. I agree that it is about suitability.
--Chris
Christopher Harrington, CISSP
Chief Technology Officer
nitrosecurity
o: 603.570.3931
c: 603.969.0592
e: charrington@nitrosecurity.com
w: www.nitrosecurity.com
Skype: chrisharrington
