[Novalug] MTU settings

James Ewing Cottrell, III jecottrell3@comcast.net
Mon Jul 27 16:23:22 EDT 2015


Thanks! I read something similar which stated that "Throughput is Proportional to Packet Size". Obvious, it's probably more Complicated, but if you have Data to Move (and Don't We All in the Age of Big Data?). More is Better, at least up to a Point.

Cisco and Juniper would most likely be able to Sell More Equipment by supporting Larger Sizes. And they Already Do! The National Library of Medicine uses an MTU of 9000 on at least Half their 1500 machines. I am sure that Many Others Do!

Infiniband is Nice, but pretty Expensive for anything but small clusers.

JIM

P.S. Did you go to Peary HS? Class of 72 here.

----- Original Message -----
From: "Joe Klein" <jsklein@gmail.com>
To: "Bryan J Smith" <b.j.smith@ieee.org>
Cc: "James Ewing Cottrell, III" <jecottrell3@comcast.net>, "Peter Larsen" <peter@peterlarsen.org>, "Novalug" <novalug@firemountain.net>
Sent: Monday, July 27, 2015 4:09:23 PM
Subject: Re: [Novalug] MTU settings

In 1999, a engineering from SPAWAR did a study about optimal MTU for each
layer 2 speed, and submitted it to IEEE. As an example, the optimal for 1
Gig was MTU 50,000,000 . The benefit was not only optimal end to end
communications, but lower the power requirements for processing each
packet, and improving the security of devices by migrating packet
fragmentation bypass by firewalls and IDS's. Sadly the document is now off
line, otherwise I would provide a link.

Many companies such as Cisco and Juniper would take a financial hit to
upgrade their products, so IEEE has move very slowly on implementing the 9k
MTU. The justification has been that an MTU upgrade would require retooling
of  chipsets and software products to make use of larger MTU.

I personally recommend using Infiniband between internal systems systems to
optimize the MTU, and
IPv6 as your layer 3 protocol. The reason is three fold for IPv6. First
IPv6 support jumbo grams, MTU's with 2^32 octets.  Second it is built into
all operating systems. And lastly, if you need to communicate to devices on
network, the Path MTU Discovery (PMTUD) works well, unlike on IPv4.

Joe Klein
"Inveniam viam aut faciam"

On Mon, Jul 27, 2015 at 3:45 PM, Bryan J Smith via Novalug <
novalug@firemountain.net> wrote:

> On Mon, Jul 27, 2015 at 3:20 PM, James Ewing Cottrell, III via Novalug
> <novalug@firemountain.net> wrote:
> > Well, all that is Very Well and Good, but the bigger question remains....
> > WHY are we STILL using 1500 byte Frames?!?
>
> Because dynamic VLAN membership based on frame size hasn't been well
> adopted, let alone well implemented.
>
> > That was the Number we used back in the 10Mbps days;
> > we are at 1,000 times that now!
> > We should but At Least using an MTU of 9000, and probably
> > closer to 64K as the Default.
>
> Hence why I've done Infiniband at some organizations.  ;)
>
> At the same time, a lot of organizations also use four (4) NICs ... 2
> bonded for a 1500 byte VLAN, and 2 bonded jumbo VLAN.
>
> > Peter, here is a Good Place for Red Hat to Lead the Way!
>
> In what regard?
>
> This is an infrastructure issue.  Red Hat _fully_ supports jumbo
> frames _very_ well.  First-hand experience (bias) talking here.
>
> In all seriousness ... I don't know if it's going to be solved
> _before_ the software is the infrastructure.  Because hardware
> compatibility rules supreme.
>
> Otherwise we'd all be on an Alpha instruction set.  ;)
>
> -- bjs
> **********************************************************************
> The Novalug mailing list is hosted by firemountain.net.
>
> To unsubscribe or change delivery options:
> http://www.firemountain.net/mailman/listinfo/novalug
>



More information about the Novalug mailing list