[c-nsp] 100 meg throughput

[Richard A Steenbergen]

On Thu, Dec 15, 2005 at 07:46:57AM +0100, Mikael Abrahamsson wrote:
> On Wed, 14 Dec 2005, Chris Cappuccio wrote:
> 
> > If your 5 minute average shows that your pipe is at 80% utilization, 
> > then it's likely that you are actually spiking at 100% at which times 
> > you are dropping packets and providing a lower quality service.
> 
> It's also interesting to know what the utilization is at what level.
> 
> It's quite possible that with imix you cannot get more than 80 meg/s of IP 
> thru a 100 meg ethernet connection, if you calculate the ethernet 
> overhead. With small packets you get a lot less than that of IP.
> 
> So 100% ethernet utilization can be 50% of linkspeed of IP packets. I've 
> had this discussion with customers before, they see 55 megs of usage on 
> their MRTG graph, we sold them 60, they're complaining, we're saying they 
> have what they bought (60 megs of ethernetencapsulated IP). This needs to 
> be specified in the contract.

Ethernet overhead guarantees that you will burn a minimum of 84 bytes per 
packet of "wire time", regardless of size.

Preamble and SFD      8 bytes
Ethernet Header    + 14 bytes
Payload + Padding  + 46 bytes
Frame Checksum     +  4 bytes
Inter Frame Gap    + 12 bytes
                     --
                     84 bytes

The term "imix" is a bit confusing, especially for calculating overhead. 
Generally speaking, most traffic is composed of lots of 1500 byte packets 
with content, being acknowledged by 44-48 byte acks, plus a bit of misc 
junk like SYNs, FINs, L7 headers, etc. On a content heavy network, you may 
see 90% of your outbound traffic as 1500, and 90% of your inbound as 44-48 
acks, so "imix" isn't particularly useful in that calculation.

Now, correct me if I'm wrong here (I'm not much of a Cisco user), but I 
believe that the data in "show int" is inclusive of the ethernet overhead 
(or at least the stuff that actually has data, like the ethernet header). 
I know that the SNMP spec calls if the standard OIDs (if*Octets etc) to 
include L2 header overhead, and the only vendor I know of offhand who 
breaks this is Juniper. Juniper strips the L2 headers at the PIC/FPC level 
before the packets reach the switch board, so classic Juniper architecture 
(pre Quad-Price PIC) is incapable of doing L2 usage statistics, mac 
filtering, accounting, etc. I think Foundry has the best CLI output for 
understanding L2 overhead, with a percentage display which takes it all 
into account:

30 second input rate: 157964176 bits/sec, 18659 packets/sec, 15.98% utilization 
30 second output rate: 133449104 bits/sec, 99647 packets/sec, 14.83% utilization

-- 
Richard A Steenbergen <ras at e-gerbil.net>       http://www.e-gerbil.net/ras
GPG Key ID: 0xF8B12CBC (7535 7F59 8204 ED1F CC1C 53AF 4C41 5ECA F8B1 2CBC)