[c-nsp] Packet fragmentation methods for QoS

[Everton da Silva Marques]

On Thu, Mar 10, 2005 at 12:12:34PM -0600, MADMAN wrote:
> 
> Everton da Silva Marques wrote:
> > Has anyone performed testing with
> > reducing MTU size at WAN links in
> > order to make the edge routers
> > (both PE and CE) to break down
> > large datagrams? How does it
> > compare to LFI?
> 
>   No since it wouldn't make much sense.  By
> simply reducing the MTU your  only making the
> transmission of packets more inefficient but
> doing nothing to gaurentee voice quality.
> With LFI you bust up the large datagrams and
> interject voice packets inbetween this newly
> busted up datagram to preserve the <20ms
> serialization that is recommended.

Please consider the following scenario, in
which voice+data traffic flows from H1
towards H2:

        W1                      W2
H1--CE1----PE1--[backbone]--PE2----CE2--H2

H: host
CE: customer edge router
PE: provider edge router
W: low speed WAN link

My understanding is:

1. MLPPP LFI over W links would break down large
   packets on CE1, insert voice packets inbetween,
   transmit them over W1, then reassemble
   large packets at PE2. The same interleaving
   would happen at W2, then reassembly at CE2.
   -- Known problems with MLPPP LFI:
   a) Low number of MLPPP bundles per PE (7507).
   b) Disassembling/reassembling work is imposed
      on PEs. If PE1=PE2 happens to be, that's
      double work at the same PE. Processing
      power of both CEs is consumed by MLPPP LFI.
   c) MLPPP bugs may affect IOS stability.

2. Low MTU on W links would make the CE1 to
   break down large datagrams. If this
   fragmentation takes place before the
   QoS queues at CE1, the voice packets
   would also have the chance to be prioritized
   inbetween the data fragments; otherwise,
   voice packets would not gain QoS benefit on
   this first W1 link, but QoS policy at W2
   would honor the preference of voice packets
   over fragmented data packets.
   -- Possible advantages of low MTU for
   voice traffic over slow WAN links:
   a) Packet fragmentation would happen only
      once, at the ingress CE, saving PE/CE
      resources.
   b) Packet reassembly would happen only
      once, at the destination host, saving
      PE/CE resources.
   -- Possible problems:
   a) Fragmentation would increase the packet
      rate over the backbone. This ought to be
      a minor effect, as such low MTU method
      should be used only for slow WAN links
      carrying voice+data under QoS; increase
      of packet rate due to fragmentation on
      these slow WAN links is expected to be
      small. Has anyone verified this?
   b) If fragmentation takes place after the
      QoS queuing at the ingress CE, voice
      packets would miss the opportunity to
      bypass lower priority data packets on
      the first W1 link. Can anyone point how
      fragmentation and QoS interact on Cisco
      devices?
   c) A given application could refuse to
      work with fragmented packet. This
      seems unlikely as most data applications
      nowdays are expected to rely on the
      underlying operating system for network
      layer packet reassembly. An issue to
      be aware of, though.

Thanks,
Everton