Re: [[e-mail]: New Version Notification for draft-jones-6man-historic-rfc2675-00.txt]

[Gyan Mishra <[e-mail]>]

Sorry accidentally hit send .. 


6MAN

Here is my take on jumbograms as I have done extensive research, testing and deployment of Jumbo frames within enterprises as well as super jumbo.

Jumbo = any frame size greater then 1500 = 1501 to 8999
Super Jumbo = any frame sizes greater then 9000 = 9000 to 65535
Jumbograms 16 bit to 32 bit field change 65536 to roughly 4.2Billion bytes

So the two manor benefits of Jumbo is cpu processing and throughout gains as the frame size increases the  IPv4 or IPv6 overhead is reduced thus increasing then bandwidth throughput.  

The basic concept I have used in deployments of Jumbo and Super Jumbo is that the network L3 or L2 port has to have a higher MTU then the hosts to avoid fragmentation which is the killer of network as a for IPv4 and ICMP unreachable is sent Type 3 code 4 if DF bit is set and the hosts have to dynamically adjust the MTU via path mtu discovery as long as fully enabled end to end along network path which we have seen issue with firewalls not passing pmtud.  From a deployment perspective of super jumbo to edge hosts the hosts have been data center servers primarily for Mainframe to Mainframe and storage backup on super jumbo 9100 over a Cisco mpls core which started as 9216 and migrated to 65535 mpls core for future bump of sever to server jumbo to slightly less 200 or so bytes less some round number like 65000 when server and mainframe hosts start supporting super jumbo up to max 65535.

The problem with fragmentation from a network perspective is that that it had to be processed in software and thus leads to increased cpu processing and significant impact as well as not being able to rely on pmtud as pmtud is generally not passed by firewalls you really have to be careful how you deploy jumbo and it rally requires one basic concept and that is that the network mtu must be greater then or equal to host port mtu to avoid fragmentation.  That is key to jumbo deployments.

So another very important concept related to Super jumbo deployment that can be applied to IPv6 jumbograms deployment is that when going from super jumbo to non jumbo endpoints.

Also to use adjust-mss going from jumbo to non jumbo endpoints on the router to avoid fragmentation.

This is a lot to digest.  Let me work on getting this all put together more cleanly.

I can work on providing some input into this draft  learnings from super jumbo deployments that we can apply to jumbograms.

Gyan


Sent from my iPhone

> On May 17, 2019, at 7:57 PM, Gyan Mishra <[e-mail]> wrote:
> 
> 6MAN
> 
> Here is my take on jumbograms as I have done extensive research, testing and deployment of Jumbo frames within enterprises as well as super jumbo.
> 
> Jumbo = any frame size greater then 1500 = 1501 to 8999
> Super Jumbo = any frame sizes greater then 9000 = 9000 to 65535
> Jumbograms 16 bit to 32 bit field change 65536 to roughly 4.2Billion bytes
> 
> So really the two manor benefits of Jumbo is cpu processing and throughout gains as you in