[tforum] RE: [it-managers] Raising the Internet MTU

Julio Cesar Facelli Julio.Facelli@utah.edu
Thu, 19 Apr 2007 09:09:40 -0600


I think that the passion comes from the research community...  

On the other hand, we have proved that jumbo frames certainly have their
place. Jumbo frames belong in environments where bandwidth and serialization
delays are low enough that intermixing jumbo frames and standard frames are
of little concern.  Also if you are doing BGP updates for large tables in a
real time environment jumbo frames are a must.

The real question is - what happens when all applications and hosts (ie
iTunes, eMule and other bandwidth hogs) are using jumbo frames?  Does our
delay and jitter on a Gigabit or TenGigabit network go back to the days of
fT1 and suboptimal frame relay speeds?  Studies like this one make good


... but I think the argument will continue until we see a working real life
implementation of jumbo frames from application to application on a large

Joshua Loveless CCIET #16638
Utah Education Network - Network Engineering
Phone: (801) 585-1133
Cell: (801) 647-4436
Email: josh@uen.org

-----Original Message-----
From: Tom Ammon [mailto:tom.ammon@utah.edu] 
Sent: Wednesday, April 18, 2007 4:38 PM
To: Josh Loveless
Cc: Joe Breen; network-research@lists.utah.edu; tforum@uen.org;
Subject: Re: [it-managers] Raising the Internet MTU

Just as a general (nontechnical) question-

It seems that the discussion over jumbo frames is almost a religious 
one- people seem to be very passionate about it, one way or the other. 
What caused this kind of polarity? Did somebody in the IETF got his 
lunch eaten over it or something?


Josh Loveless wrote:
> I do not have any hard data on actual jitter caused by jumbo frames.
> However, you might be interested to know the UEN applies a 9000 byte MTU
> any Gigabit Link in our PoPs.  In most cases our maximum jitter is .1 ms.
> We have been running this way for about three years now. 
> We have found in most cases the actual packet size is limited by the
> applications sending and receiving the data.  Also keep in mind that the
> at the edge where the applications connect is usually a standard MTU.  We
> use the jumbo frames primarily for routing protocol updates.  Using a
> frame allows BGP to converge significantly faster.
> We also require all new ethernet connections (with a few exceptions for
> rural areas) to support jumbo frames, just as an FYI.
> Thanks, 
> Joshua Loveless CCIET #16638
> Utah Education Network - Network Engineering
> Phone: (801) 585-1133
> Cell: (801) 647-4436
> Email: josh@uen.org
> -----Original Message-----
> From: Joe Breen [mailto:Joe.Breen@utah.edu] 
> Sent: Wednesday, April 18, 2007 11:47 AM
> To: Tom Ammon
> Cc: network-research@lists.utah.edu; tforum@uen.org;
> it-managers@lists.utah.edu
> Subject: Re: [it-managers] Raising the Internet MTU
> Tom Ammon wrote:
>> Has anyone seen any data concerning large frame size's impact on jitter?
> Tom,
> As the speed of the link goes up, the impact of larger frame sizes goes 
> down unless the frame sizes grow significantly.  What you bring up is a 
> valid question, though, because you don't want latency sensitive 
> volkswagen beetles stuck behind huge semi-trucks, or latency sensitive 
> voice packets trapped in the queues behind large file transfers.  Some 
> discussion exists in some of the links off of the main link I sent out. 
>   For example, Phil Dykstra's discussion:
> http://sd.wareonearth.com/~phil/jumbo.html
> I will have to dig up some other links I have seen on the subject. 
> There have been some rigorous papers dealing with the question.
> In practice, most jitter on voice and multimedia comes from dropped 
> packets, varying latencies due to changing paths, routers running under 
> high utilization, strange asymmetric routes or host issues.  As long as 
> the pipes are fast, packet drop negligible and the routers are truly 
> running wire speed, jitter issues tend to be lower.   Buffer sizes on 
> hosts, routers and switches can affect jitter.  Host CPU utilization 
> also affects jitter.  How well a host is handling interrupts also 
> affects jitter.  Many factors affect jitter, especially in the Wide Area 
> Network.
> Large frame sizes are typically not common on the internet due to ISPs 
> and end-customer devices not supporting Jumbo frames.   However, in the 
> research world, the national and international backbones do support 
> jumbo frames as well as very high speed pipes.  Because the pipes are 
> now 2.5Gigabit to 10Gigabit and the max Jumbo frame is only 9kbytes.  A 
> 9kbyte frame on a 10Gig link is actually less of an impact than the 
> standard 1500byte frame on a 10Megabit link.
> 	--Joe
>> Joe Breen wrote:
>>> For those following network performance issues,
>>> As networks have progressed in speed from 10Megabit to 10Gigabit, the 
>>> size of the transmission unit has stayed the same.  This fact impacts 
>>> the performance of networks.  The following link is a very useful 
>>> starting point on Jumbo-frames and larger Maximum Transmission Units.
>>> http://www.psc.edu/~mathis/MTU/
>>> Matt Mathis of the Pittsburgh SuperComputing Center is one of the 
>>> leading investigators into larger packet sizes and their impact on 
>>> performance.  He is also one of the co-authors of a new Request for 
>>> Comment (RFC) regarding path MTU discovery, the ability for end 
>>> machines to discover the maximum transmission unit.
>>>     --Joe

Tom Ammon
Network Engineer
M: (801)674-9273

Center for High Performance Computing
University of Utah