🚀 go-pugleaf

In article <19971104200217483265@ppp14.dial-in.netguide.dk>, 
pope@my.terminal.dk says...
>
>Robert Wilkens <rwilkens@ccur.com> wrote:
>
>> elel@iexist.flw.lucent.com (Eric L. Edberg) writes:
>> >     Changing MTU and MSS will NOT affect throughput unless you
>> >        have high CRC errors on the down or up stream paths.
>> 
>> This is just not true.
>> 
>> If you have your MTU set too small, packets will have to be fragmented.
>> 
>> Among other things, this means you'll probably duplicate or triplicate
>> header information for each packet.
>

true, you always want the max possible MTU to be negotiated

>All the servers on ethernet you are on are using the 1500 bytes MTU
>(maximum transmission unit), so it will send packets this size all the
>way to your server (as far as I remember, I can not remember if it is
>negotiated at start of TCP sessions though). 

You are thinking MTU path discovery, IP does this, not TCP. TCP(slow start 
algorithm?) negotiates window size, how many IP packets I'll hear before I 
send an acknoweldgment.

>So the server that will
>have to fragment the packages will get extra loaded.

some CPU overhead, is it really that bad? I doubt it. CPU's are screaming 
these days, this isn't the issue, the issue is protocol overhead that is 
driving up your network utilization and chewing up bandwidth on headers and 
not your data.

>
>And remember, as TCP/IP packet header is large. 

40 bytes acutally, 20 bytes for TCP and at LEAST 20 more for the IP
header, more if there are options flags set, padded  if not enough.

>An ICMP packet is 8 bytes

probably, check rfc 792.



>bytes. And TCP is 12 bytes (or is it 16 ?). So they make up about 1% of
>all traffic.
>
>BTW: A good ethernet card will do about 1100-1150 kb/sec using TCP/IP.
>The rest up to the theoretical 1250 kb/sec (10Mbit = 8 bits * 1250kb) is
>mostly protocol overhead. So on ethernet 10% is actually wasted.


It's not dedicated bandwidth, it is a shared broadcast medium, and when the 
you approach the supposed ~37% max utilization you will begin to feel a severe 
performance hit, I have witnessed this time and again on corporate net's so am 
a believer. I'm willing to bet very few ppl are seeing this kind of 
utilization on there segments. I can leave an analyzer running for hours at a 
time and not fill a 4 meg packet buffer, so long as I'm not doing anything 
myself that is.

My average speed to my gateway is nearly 500K consistently when doing adhoc 
testing. once it passes my gateway, bounces off two or three nodes and hits 
the Net at large, my throughput is more like 1.5KBps to 50KBps.
As someone else noted though, testing a live network is utterly pointless.


Sean Cheney	
Network Systems Engineer
International Network Services