Re: [c-nsp] Handling Out of order packet

2011-04-04 Thread Justin M. Streiner

On Mon, 4 Apr 2011, arulgobinath emmanuel wrote:


When trying to load balance through 2 Ethernet paths  (jitter 4ms - 10ms and
delay 10ms)  using per packet load  sharing severely impact the throughput.
(eg per path 90Mbps but when enabling both paths 30mbps ) . I suspect the
issue due to the out of order packet. Is there any way similar to MLPPP that
can handle the out of order packet in Ethernet environment.


You're probably going to need to use per-flow load-sharing.  The 
possibility of having out-of-order packets is one of the downfalls of

per-packet load-sharing.

jms
___
cisco-nsp mailing list  cisco-nsp@puck.nether.net
https://puck.nether.net/mailman/listinfo/cisco-nsp
archive at http://puck.nether.net/pipermail/cisco-nsp/


Re: [c-nsp] Handling Out of order packet

2011-04-04 Thread Jared Mauch
Is there a reason you are not just doing per-flow?  Have you looked at running 
LACP?  Doing per-packet as you mention sometimes causes problems with the end 
hosts if they have a poorly behaving IP stack, or low buffering.  It also 
creates extra work for the hosts.

You can change the hashing system for LACP to perform differently based on your 
requirements (and platform).  Knowing more about your setup will help provide a 
better recommendation.

I would leave per-packet as a last resort in almost any environment.

- Jared

On Apr 4, 2011, at 10:32 AM, arulgobinath emmanuel arulg...@gmail.com wrote:

 HI,
 When trying to load balance through 2 Ethernet paths  (jitter 4ms - 10ms and
 delay 10ms)  using per packet load  sharing severely impact the throughput.
 (eg per path 90Mbps but when enabling both paths 30mbps ) . I suspect the
 issue due to the out of order packet. Is there any way similar to MLPPP that
 can handle the out of order packet in Ethernet environment.
 
 Thanks in advance.
 Gobinath.
 ___
 cisco-nsp mailing list  cisco-nsp@puck.nether.net
 https://puck.nether.net/mailman/listinfo/cisco-nsp
 archive at http://puck.nether.net/pipermail/cisco-nsp/

___
cisco-nsp mailing list  cisco-nsp@puck.nether.net
https://puck.nether.net/mailman/listinfo/cisco-nsp
archive at http://puck.nether.net/pipermail/cisco-nsp/


Re: [c-nsp] Handling Out of order packet

2011-04-04 Thread Nick Hilliard

On 04/04/2011 15:42, Jared Mauch wrote:

Is there a reason you are not just doing per-flow?  Have you looked at
running LACP?


LACP is orthogonal to the hashing algorithm.


I would leave per-packet as a last resort in almost any environment.


I have seen per-packet cause performance gain, but only in extremely 
distressed circumstances.  You're right that it should generally be 
disabled, because it causes all sorts of trouble.


Nick
___
cisco-nsp mailing list  cisco-nsp@puck.nether.net
https://puck.nether.net/mailman/listinfo/cisco-nsp
archive at http://puck.nether.net/pipermail/cisco-nsp/


Re: [c-nsp] Handling Out of order packet

2011-04-04 Thread Jared Mauch
See below

Jared Mauch

On Apr 4, 2011, at 4:27 PM, Nick Hilliard n...@foobar.org wrote:

 On 04/04/2011 15:42, Jared Mauch wrote:
 Is there a reason you are not just doing per-flow?  Have you looked at
 running LACP?
 
 LACP is orthogonal to the hashing algorithm.

The hardware hashing does per flow only, so it is related :-) I've never seen 
hardware that does anything but at least. 


 
 I would leave per-packet as a last resort in almost any environment.
 
 I have seen per-packet cause performance gain, but only in extremely 
 distressed circumstances.  You're right that it should generally be disabled, 
 because it causes all sorts of trouble.
 

True, but the topology sounds like diverse paths being used for load sharing vs 
redundancy. Lacp fast mode can help here too depending on the platform. 

Without knowing more it's hard to tell. But it sounds like an Ethernet topology 
and the 802.3 tools may help here. 


 Nick
 ___
 cisco-nsp mailing list  cisco-nsp@puck.nether.net
 https://puck.nether.net/mailman/listinfo/cisco-nsp
 archive at http://puck.nether.net/pipermail/cisco-nsp/

___
cisco-nsp mailing list  cisco-nsp@puck.nether.net
https://puck.nether.net/mailman/listinfo/cisco-nsp
archive at http://puck.nether.net/pipermail/cisco-nsp/


Re: [c-nsp] Handling Out of order packet

2011-04-04 Thread arulgobinath emmanuel
Thanks All,
the setup involved

 Fa1/0-Wireless
Bridge  ||   Wireless Bridge --- fa0/0
End HostROUTER
1
ROUTER 2   -- End Host ( Windows 7)
  fa1/1   Wireless
Bridge   || Wireless Bridge - fa0/1

end host connected to router (Cisco 2821) and hwic port connects to  two
wireless bridges same setup replicated other end. Since the wireless
physical media cause the jitter  delay .  If its pure Ethernet may be the
out of order packet rate  less.

I'm not using per flow because there are less number of hosts hence unable
to archive required throughput.  I've tried etherchannel(Pagp )but the same
issue since the traffic shared using flow based.  Afterwards i've replaced
switch with router to do the per packet load sharing .

Since the per packet load sharing can cause some out of order packet i'm
looking into  options(tunneling ? )  which provide some buffering in the
router which regulates the traffic.  Except IPS feature no other solutions
yet .

Thanks again.
Gobinath.



On Tue, Apr 5, 2011 at 3:42 AM, Jared Mauch ja...@puck.nether.net wrote:

 See below

 Jared Mauch

 On Apr 4, 2011, at 4:27 PM, Nick Hilliard n...@foobar.org wrote:

  On 04/04/2011 15:42, Jared Mauch wrote:
  Is there a reason you are not just doing per-flow?  Have you looked at
  running LACP?
 
  LACP is orthogonal to the hashing algorithm.

 The hardware hashing does per flow only, so it is related :-) I've never
 seen hardware that does anything but at least.


 
  I would leave per-packet as a last resort in almost any environment.
 
  I have seen per-packet cause performance gain, but only in extremely
 distressed circumstances.  You're right that it should generally be
 disabled, because it causes all sorts of trouble.
 

 True, but the topology sounds like diverse paths being used for load
 sharing vs redundancy. Lacp fast mode can help here too depending on the
 platform.

 Without knowing more it's hard to tell. But it sounds like an Ethernet
 topology and the 802.3 tools may help here.


  Nick
  ___
  cisco-nsp mailing list  cisco-nsp@puck.nether.net
  https://puck.nether.net/mailman/listinfo/cisco-nsp
  archive at http://puck.nether.net/pipermail/cisco-nsp/

 ___
 cisco-nsp mailing list  cisco-nsp@puck.nether.net
 https://puck.nether.net/mailman/listinfo/cisco-nsp
 archive at http://puck.nether.net/pipermail/cisco-nsp/

___
cisco-nsp mailing list  cisco-nsp@puck.nether.net
https://puck.nether.net/mailman/listinfo/cisco-nsp
archive at http://puck.nether.net/pipermail/cisco-nsp/