Comments on: How to optimize your DSL broadband performance with Jumbo Frame support

By: George Ou

George Ou — Sun, 06 Jul 2008 03:41:56 +0000

"In fact, in so far as a person uses 1464 as an MTU instead of 1500, they are losing 2.5% of their per-packet performance."

1464 is the actual payload size; 1492 is the actual MTU because it's DSL with an 8-byte PPPoE overhead. The fact that you don't understand this is simply scary.

Second, you're not losing 2.5% performance between MTU 1500 and MTU 1464. Smaller packets have lower payload to header ratio. The difference in additional overhead is less than .2% difference and if you don't understand that, go take some basic networking classes.

George Ou

By: George Nutsou

George Nutsou — Sun, 06 Jul 2008 03:34:53 +0000

Mr. Ou,

I understood what you said; what you don't understand is that it is irrelevant -- moot. In fact, in so far as a person uses 1464 as an MTU instead of 1500, they are losing 2.5% of their per-packet performance.

If your local network doesn't support 9000, then your MTU should be set to 1500 (or the limit of your switch, if it isn't 1500 or 9000) and forget about it. No need for multiples of anything. IP does it for you, as long as your network isn't screwed up.

The bottom line is this: The 2nd paragraph of this story is factually incorrect. A 1500 MTU does NOT cause twice as many packets over a 1492 link. IP+ICMP notes there is a link with a smaller MTU in the path, and lowers the MTU on that path accordingly. Result: no fragments.

By: George Ou

George Ou — Fri, 27 Jun 2008 05:43:37 +0000

Mr. Nutsou, seems you haven’t tested anything because you wouldn’t be talking the way you are if you did. MTU 9000 isn’t supported by many jumbo-frame enabled switches and it isn’t supported by many Ethernet adapters. What I told users to do is to use clean multiples of 1464 and use the largest MTU their equipment supports by TESTING with do-not-fragment flag. It seems like you don’t even understand the illustrations I put up.

George Ou

By: George Nutsou

George Nutsou — Fri, 27 Jun 2008 05:38:19 +0000

You don’t seem to understand IP, ICMP, and the Don’t Fragment flag. Something is broken in your network if you are fragmenting your packets. Otherwise, your packets destined to/from the Internet will automatically take the size of the smallest MTU in the chain to the remote server, likely the 1492 of your PPPoE. All the time spent figuring out multiples of the hypotenuse are wasted.

Instead, just pick the largest jumbo MTU that works with all your GigE devices and GigE switches/routers. 9000 is frequently used, although not the limit.

By: George Ou

George Ou — Sat, 07 Jun 2008 04:45:25 +0000

MTU discovery isn’t a sure thing, though the newer OSes should be more reliable. To make it always work correctly, you use clean multiples for your MTU and it will always result in cleanly and evenly chopped packets.

By: George Ou

George Ou — Sat, 07 Jun 2008 01:08:53 +0000

It’s not an OS improvement, it’s broadband speed improvement. You’re effectively sending less overhead and fewer packets per second over the broadband link. It’s probably nothing earth shattering but it is better.

By: Alok

Alok — Sat, 07 Jun 2008 01:07:20 +0000

What and how much improvement will be seen by performing this? I have DSL, using Vista.

By: Wes Felter

Wes Felter — Fri, 06 Jun 2008 23:56:01 +0000

TCP should automatically detect the proper MTU; is that not happening?

Also, AFAIK 100Mbps Ethernet does not support jumbo frames, so what happens when a jumbo frame gets to your router or modem?