On 2021-08-17 14:23, Justin Parisi wrote:

Referrals probably have the same issue. Will need to test when I get back into the office.

Unless you seek to eliminate/avoid, all CI (Cluster Interconnect) traffic at all cost, I don't really see that that fact matters too much [at mount time]. (Good point about FlexGroups though... I don't have a good solid solution for those scenarios)

We have since long (years) strived to avoid CI traffic as much as we can, but we're not as diligent anymore. Note: we do *not* use that DNS function ONTAP has in it which is intended to help spreading all the traffic (mount from clients) to various LIFs (based on the "load" on the nblades). Again, this takes place at mount time, once a mounts sits on a LIF it stays there...

That DNS function (it is based on an algorithm using various "weights") suffers from serious drawbacks in some situations and we can't take that risk in our environment. Anyone who has seen and experienced a "mount storm" will know what I mean.

How important locality is (minimal CI traffic) for you I guess can depend a bit on what model/power of CI switch you have. If there's enough bandwidth for the workload streams going across it, it only induces a little extra latency -- that may matter for some workloads, but I'd argue that for most it won't.

Once you're on CI switch with 100 GbE, my take is that most people don't have to bother that much about it. With 100 GbE the latency is lower as well

/M

Justin Parisi wrote:

But referrals only help with initial mount; after the mount is established, the connections stay where they were established. If a volume moves to a new node, that doesn't keep locality. They're also not much help with FlexGroup volumes.