[f-nsp] BigIron 15000 load balancing

Niels Bakker niels=foundry-nsp at bakker.net
Fri Oct 6 09:07:54 EDT 2006


* fahad.alikhan at gmail.com (FAHAD ALI KHAN) [Fri 06 Oct 2006, 05:50 CEST]:
>We have installed BigIron 15000 in our network. My question is does it 
>support per packet load balancing as Cisco does or it has per packet 
>load balancing as Juniper.
>
>As im seeing, it is doing network based load balanced by default which 
>is a per flow actually. Does any one help me in this regard.

You fail to mention some useful information, such as whether you have 
IronCore or JetCore, whether you're using 10GE or not, if you're trying 
to build aggregated links or do equal-cost multipathing.

Per-packet load-balancing can lead to packet reordering, which is 
generally considered a bad thing.

Juniper doesn't do per-packet, what is called per-packet is actually 
per-flow on IPII and up (and you are in trouble if you still try to do 
something useful with an IP 1).

JetCore supports switch and server trunks (aggregated links, 
Port-channel in Cisco talk).  The algorithm for the former looks at the 
destination MAC address of each frame and puts it on a member port in 
the trunk based on a hash of it; a server trunk hashes source and 
destination MAC addresses, and depending on software release also on 
source and destination IP address and TCP/UDP port numbers. (Multicast 
is always sent down the primary port.)

I can't think of a case where a switch trunk is preferable over a server 
trunk.

I understand that for the RX platform the frame is XORed and then sent 
down the appropriate modulo port; thus you can have a good distribution 
over aggregated links with an amount of member ports that is not a power 
of two.


	-- Niels.

-- 



More information about the foundry-nsp mailing list