Re: pty(4) 1024 bytes buffer limit

[Thor Lancelot Simon <tls%panix.com@localhost>]

On Fri, Sep 09, 2011 at 05:11:17PM -0700, Erik Fair wrote:
> 
> ... unless they're using jumbo frames. Potentially 9Kbytes, depending upon 
> NICs and switches.

...which reminds me that someday, I will eventually finish my "9k MTU
demonstrated harmful" informational RFC.

The original research which seemed to show a benefit for 9K MTU was done
with NFS over UDP (so IP layer fragmentation instead of TCP layer
segmentation) on systems with a 16K page size and a page-based memory
coherency algorithm for multiple processors.  Guess what the NFS RPC
size limit was set to?  Right, 8K.  If they'd used 16K there, they would
have concluded they needed a 17K MTU.

In practice, large MTU helps, particularly for receive (modern adapters
do TCP segmentation offload on send so you get the whole efficiency
benefit of a large MTU for your stack, and then some), but 9K is a very
bad choice of size: on most systems it means you allocate 4K three times
and waste the last 3K of it.  The FDDI MTU of 4K would have been a much
better choice; for some applications 8K-plus-headers is good too.

Thor