[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Reopening jumbo frames in IS-IS

Curtis Villamizar wrote: 
In message <42BF0568.5090708 at tony.li>
Tony Li writes:




Jumbo frames exists only to avoid the performance drop that would
otherwise potentially cripple core networks using GbE or 10GbE and
aggregating higher MTU links.




Another, and perhaps more common, reason for jumbo frames is to support
high speed data center and storage applications.  Many modern systems
have an 8KB primary memory page size, and can be most efficient at bulk
transfers when they can make use of the page as a transfer unit.

Tony 



Excluding NFS and a few other less used protocols there is little
reason for end system jumbo frames.


There is lots of practical evidence that this isn't true; it's one of
the reasons that Infiniband is very successful in environments such
as high performance redundant database clusters, for example.

Not that this is relevant to the IS-IS issue.

It is widely understood that NFS
8K UDP packets are a bad design.


That is irrelevant to the IS-IS issue.

IMHO endorsing jumbo frames in end systems goes to much against what
IEEE would like to see and does so for the benefit of protocols that
should have a new revision that eliminates the need by being more MTU
aware.  This also can improve their performance radically when used
over a WAN or VPN.


But the typical use of jumbo frames is in highly tuned LAN-based clusters.

Not that this is relevant to the IS-IS issue.

Also, we're not trying to rewrite the IEEE standard; we're trying to deal
with practical reality in the data center.

   Brian