[PLUG] Cannot set mtu in profile.d on Ubuntu 8.10
bain at devslashzero.com
Mon Apr 19 03:26:21 PDT 2010
> On Tue, Apr 13, 2010 at 12:39 PM, Abhijit Bhopatkar
> <bain at devslashzero.com>wrote:
>> The cvs will have a huge bulk traffic going in and out. The main
>> problem with MTU is not at server or client side, but is usually in
>> the middle, the routers. The problem is that some routers on the
>> network are not able to handle large packets very well. They start
>> building up huge chains of backlog packets and eventually cause TCP
>> retransmissions. The problem keeps aggravating very quickly and within
>> seconds the network traffic for that particular connection starts
>> grounding to halt. Addition of VPN on top of CVS might just be tipping
>> the sizes beyond the trigger point for your network setup. The problem
>> is typically intense when a cheap router is being used to share a
>> broadband connection on LAN.
> In the context of what you mentioned about the chain of backlog packets
> building up at the router:
> If a router cannot handle the size of a packet it receives, will it drop the
> packet, or will it try to fragment it?
> If I understand the above passage correctly, what you are suggesting is, if
> a router cannot handle large packets, it is going to cause re-transmissions
> and very soon the network will become sluggish. In this case will that
> router become a bottleneck for the entire LAN connected through it? The
> reason I ask this is because, if I am able to determine that the router is
> not becoming a bottleneck for the entire network, then would it be safe to
> say that the router is probably not responsible for mis-handling large
Ok i must first confess i do not understand the logic behind it fully.
But no, the router does not become bottleneck for entire LAN in the
situations I was pointing at.
Each router, especially the ones that do NATing have chains of packets
built for each of client they are forwarding to/from. Now what
happens, if you send large number of packets which are larger in size
than the packet size allowed on the other side of the router, is that,
the router will do the fragmentation itself. This fragmentation along
with some problems with TCP congestion control algorithms (that i
don't understand), make the backlog packet chain for that particular
connection start going beyond what router can handle and eventually
the queue overflows causing retransmissions, but since the original
pattern does not die down the retransmissions also create the same
situations and eventually the particular connection dies down. Now a
good router based on real time OS won't probably let other queues and
connections get affected.
Again this is all speculation created by this weird brain of mine
based on some literature found on internet (you do know that if its on
the web it must be true, right? ;) ) so do take it with many pinches
of salt. Ultimately lowering the MTU also fixed a similar problem i
was having with my DSL connection few years back.
More information about the plug-mail