On Thu, 2 Aug 2001, David Sinn wrote:
> By default CEF, and fast/optimum switching does per-flow load-balancing.
> You have to explicitly turn on per-packet load-balancing with the "ip
> load-sharing per-packet" command. Regardless of what Cisco, or the
> documentation call's it, by default it is per-flow (actually
> per-destination which means that a flow will follow the same link).
> Engine 2 LC's on GSR can not do per-packet load-balancing, regardless of
> whether you can enter the command or not.
Actually it is per-flow, since CEF hashes on the source as well as the
destination.
-Scott
>
> David
>
>
> -----Original Message-----
> From: Kinczli Zoltán [mailto:Zoltan.Kinczli@Synergon.hu]
> Sent: Wednesday, August 01, 2001 7:49 AM
> To: David Sinn; Fillet Platoon
> Cc: cisco-nsp@puck.nether.net
> Subject: RE: PBR
>
>
> David,
>
> Most of our customers do use CEF. Many of them are using
> it because the per-packet load-sharing feature. (Others do have
> really huge routing tables, where other route caches are not
> really effective.) They tipically share the load between 4 2Mbps lines.
> (They could as well use more links...)
>
> According to the i/f counters the pps values, and the load,
> calculated over 30 sec, are almost exactly the same on the links.
>
> Ok, not all the packets are of equal size, so you are right
> to tell us, it's not a strict load sharing algorithm. The load
> on the link is the number of bytes transmitted, not the number of
> packets, perfectly true.
>
> An algorithm, you desire, would be such complex one, which would
> - i guess - cause the traffic to be process switched. Are you sure
> it would worth the effort/overhead/etc??
>
> --Zoltan
>
> PS: and i'm sorry to argue, but a given flow it not using the same
> link out of a given router, or else it would not be called per-packet
> load sharing
>
>
> -----Eredeti üzenet-----
> Feladó: David Sinn [mailto:dsinn@microsoft.com]
> Küldve: 2001. augusztus 1. 3:35
> Címzett: Fillet Platoon
> Másolatot kap: cisco-nsp@puck.nether.net
> Tárgy: RE: PBR
>
>
> The simple answer is "no", but you have other options.
>
> Cisco's load balancing is not intelligent, and is not balancing. It is
> a fixed hash over the given number of links that cause a given flow to
> always use the same link out of a given router. There is no feedback
> mechanism. Thus you can have very clumpy traffic (especially on slower
> speed links).
>
> If you aren't on a GSR, then you could turn on CEF per-packet load
> balancing but this means you need to run CEF, and very few people do
> this, so you mileage will vary (i.e.. If you don't want to run CEF you
> can turn off all route-caching and the router will per-packet load
> balance by default. This method is not very elegant, but depending on
> load MIGHT be acceptable. Either way you are telling the router to load
> balance the traffic on a per-packet basis across the parallel links.
> This can result in out-of-order delivery of traffic, so consider if that
> is important to you. Also it is not a guarantee that the traffic will
> still be well balanced, but is a better shot then having the flow based
> load balancing you have now and resulting in the clumpy traffic you are
> seeing.
>
> You can also consider more L2 load balancing such as MLPPP or ATM-IMA.
>
> ATM-IMA can work if you are talking about T-1's, and last I checked the
> PA/NM's supported 8 T1's. It does mean you have to run ATM, which is
> mostly considered a four letter word, so think about it deeply before
> you jump. You also loose a fair amount total throughput due to
> overhead, so you have even more to consider if you are on slow links.
>
> MLPPP could also be an option, and is possibly your best if you control
> both ends, and can take the CPU hit that it creates.
>
> All in all it depends on what box you are on, and how much pain you can
> endure.
>
> David
>
>
> -----Original Message-----
> From: Fillet Platoon [mailto:fplatoon@yahoo.com]
> Sent: Monday, July 30, 2001 8:39 PM
> To: cisco-nsp@puck.nether.net
> Subject: PBR
>
>
> hello guys,
>
> my goal is to load balanced my point-to-point
> links that have 8 parallel serial connection, so i
> used EIGRP as my routing protocol and i used MHSRP
> along with PBR on some of my selective source address.
> Is there a routing protocol that intelligent enough to
> detect that my Serial 0 connection is congested and i
> will automatically redirect it to serial 1 or Serial 2
> that is not congested? If not is there a way to solve
> it? Is there a features in Cisco IOS that
> automatically detect that the link is congested and
> automatically redirect it to another link that is not
> congested?
>
> rgds,
> fillet
>
> __________________________________________________
> Do You Yahoo!?
> Make international calls for as low as $.04/minute with Yahoo! Messenger
> http://phonecard.yahoo.com/
>
This archive was generated by hypermail 2b29 : Sun Aug 04 2002 - 04:12:48 EDT