Can anyone explain to me, the LAG sharing balancing hash algorithm XOR,CRC-16,CRC-32?


Hi all, I have a problem running a bert test through a LAG LACP between a Huawei Switch and our Extreme X460G2. The problem is that the Extreme is not making load balancing, I suppose because the two Exfo Machines that run Bert test use the same port during all the test so any address based balancing is useful forr the test. So I want to test this other balancing types Xor, CRC-16 and CRC-32, but I want to undestand how it works, maybe you know about another way to balance the load in the LAG LACP. The Huawei switch is making good balancing , my ISP L2 provider told me that hes using packet balancing, the first packet for one port, the second for the other, the third equal as first, and that repeatively, but I coudnt find something similar in Exteme

Thanks in advance

4 replies

Userlevel 6
Hi. Packet balancing is not good. It will create mis-order in the flow. You can chose different hash algorithms. L2, L3, L4. In any of those algorithms if there is only one flow, it will not be balannced. Regards Z.
Userlevel 7
I agree with Z on the fact that round-robin is bad. This is how you end up with end systems trying to reorder packets, and usually failling miserably. If you have multiple flows, hashing algo will help you balance per-flow. You have the choice between L2, L3, L3_L4 and custom. It gives you an idea of what portion of the headers are used to differentiate the flows. Using L2, for example, between two routers would be bad. Custom is usually a good candidate as it will look at various portion of the headers, L2 to L4. Off the top of my head, I don't remember the exact bits, but I would give it a try. Using custom you also have the possibility to have larger LAG and change between xor easily. Related to LAG you might also be interested with features like resilient hashing.
Userlevel 7
Hi,

with CRC-16 or CRC-32, a cyclic redundancy checksum of the selected header fields is used instead of using a simple XOR. This can help with load sharing of artificial flows created by some testing gear (or perhaps some applications). This still balances per flow, not per packet.

Erik
Thanks you very much, so It seems that I have to find a type of test that could make many conections so the extreme would balance the traffic

Reply