is odd number of links in lag group ok

I have (2) 10 gig links bundled in a lag to my upstream internet provider.
and we need more internet capacity. Is it cool to add a third 10 gig to my
existing 20 gig lag internet connection?

I'm asking since I heard in the past something negative about odd numbers of
lag members. .but I also have heard that it's not a big deal. Let me know
please

-Aaron

Much of this depends on the hardware, software and what hashing is used inbound
outbound traffic directions, etc.

It will likely work the way you expect, but one may be warmer than the other if
traffic ends up overloading a single bucket in the hash.

- Jared

We haven't had a major issue when loading traffic over even or odd
links. If you have decent hardware and software, it should all be fine,
particularly if your traffic is all or mostly IP.

If you've got non-IP traffic in there, and your box cannot look into the
payload to determine entropy, then things could get interesting. But
this will happen even when you have even links... it's not anything
specific to how many member links you have in the LAG, but rather, the
router's need to maintain per-flow load balancing with limited
information beyond Layer 2 data.

That said, an even number of links just leaves the warm & fuzzies turned
on :-)...

Mark.

Well, power of two, surely? 6 is even, but is not a "good" number of links for a lag group.

(Extreme X8 supports up to 64 links in a group, and Enterasys up to 127 if I recall correctly. Though Enterasys, unusually, seems to support round-robin load balancing in which case the argument for powers of 2 numbers of links ceases to make sense.)

It will work fine if you have a good modern router. Consider this; all evenly grouped LAGs are odd in their failed conditions.

-Ben

This is implementation specific.

Consider hash result returning range 0-15.

Then number of port count as divisor of 16/n must result in an integer value.

To workaround this issue, you either have larger range of hash
results, in which case unequal distribution has small bias, or you
have different hash functions depending on port count so that that
hash returns appropriate range for given port count.

Certainly solvable problem, and generally non-issue.

As others have noted, there can be implementation specific issues that
you can't necessarily predict but most typically when I hear "odd vs
even" discussions, usually the caveat is not a trunk but a redundant
connection. Putting three links on router A and two links on router B
obviously doesn't work well.

While it goes without saying that you need the same (can be 5!) number of links to each router in a multichassis LAG, what isn’t so obvious are things like port groups etc.

If you have an oversubscribed platform, you might need to look at running each wire in a LAG to different port groups, and then look at things like switch ASICs and span those as well.

Even try to span diverse slots/modules if you can.

But 5 6 or 4 per chassis shouldn’t make a huge difference.

-Ben

I'm not sure the OP was talking about MC-LAG. Just a regular LAG.

Mark.