[ale] PPP Perfromance
Dow Hurst
dhurst at kennesaw.edu
Fri Oct 31 15:27:48 EST 2003
The local LAN hardware that supports your LAN has to be able to handle
jumbo frames if you want to have larger packet sizes than 1500 bytes.
That is why Linux clusters using Gigabit switches have had to pay more
for the switches that would support jumbo frames if they want to
maximize the efficiency of the bandwidth. The 1500 byte max is also a
limiting factor on PPPoE since the LAN behind the DSL router should have
an MTU of less than 1500. How much less is debatable but the Roaring
Penguin software recommended 1412 if I remember correctly. You want the
LAN ethernet frames to be efficiently packaged inside the PPP frame
which is packaged inside the ATM frame. No wonder everyone wants a
straight ethernet instead of PPPoE for their broadband!! Fragmenting
the local LAN packets just lowers the efficiency of your PPPoE
connection and the overhead of the router's rebuilding the fragmented
packets.
To switch to another topic of LAN efficiency at a different layer: For
TCP and UDP packet size over a local LAN I've played with NFS packet
sizes to see what is most efficient. Here in our VPN at KSU I found that
udp and tcp required different values. So, packet size at the ethernet
layer and at the higher layer has a big effect on efficiency. Within a
local LAN on normal hardware 1500 MTU is best since you get the largest
ethernet packet per packet transmission while a UDP/NFS packet size of
64K instead of the default 8K might increase efficiency. It is good to
spend an afternoon testing such things to tune the network traffic. I
use SGI IRIX so my NFS packet sizes for UDP and TCP might be tuned
differently than a Solaris or Linux based LAN. I would imagine that as
Gigabit switches drop in price and that the models with jumbo packet
capability also drop that eventually we might be able to take advantage
on home LANs for passing video and such. I don't know alot about jumbo
packets so I can't say anything other than just this little bit. Jeff
Layton might have a comment?
Dow
Doug McNash wrote:
>
> The max number of data bytes one can put in a 802.3 Ethernet frame is
> 1500 per IEEE standard
>
>
> On Fri, 31 Oct 2003 10:53:49 -0500
> Geoffrey <esoteric at 3times25.net> wrote:
>
>> Doug McNash wrote:
>> Getting back to the mtu issue, I'm now curious as to whether that
>> might be part of the problem why I couldn't get to my mother-in-laws
>> benefit web site. I've played with mtu a bit and found a couple
>> things out. Currently all my ethernet networks are set at mtu 1500.
>> I can easily reduce this by simply:
>>
>> ifconfig eth0 mtu 1000
>>
>> But if I attempt to increase it, I get an error. Me thinks this
>> would because there's a max set somewhere, although after running
>> around in /proc for a while, I couldn't find anything. Anyone else
>> have any ideas?
>>
>>
>> --
>> Until later, Geoffrey esoteric at 3times25.net
>>
>> Building secure systems inspite of Microsoft
>>
>> _______________________________________________
>> Ale mailing list
>> Ale at ale.org
>> http://www.ale.org/mailman/listinfo/ale
>
>
> --
> Doug McNash <dmcnash at yahoo.com>
> _______________________________________________
> Ale mailing list
> Ale at ale.org
> http://www.ale.org/mailman/listinfo/ale
>
--
__________________________________________________________
Dow Hurst Office: 770-499-3428 *
Systems Support Specialist Fax: 770-423-6744 *
1000 Chastain Rd. Bldg. 12 *
Chemistry Department SC428 Email: dhurst at kennesaw.edu *
Kennesaw State University Dow.Hurst at mindspring.com *
Kennesaw, GA 30144 *
************************************************************
This message (including any attachments) contains *
confidential information intended for a specific individual*
and purpose, and is protected by law. If you are not the *
intended recipient, you should delete this message and are *
hereby notified that any disclosure, copying, distribution *
of this message, or the taking of any action based on it, *
is strictly prohibited. *
************************************************************
More information about the Ale
mailing list