conf: add mapall upcall to xCAT group bindings

[thiell]

Adopt the new mapall upcall in the xCAT example group bindings: one lsdef -t group -i members call replaces one lsdef call per group, for both static and dynamic groups. Group names containing : are skipped and remain resolvable through map.

doc/sphinx/config.rst updated to match.

[kcgthb]

Actually, I'm not really sure about this one. 🤔

mapall works great when the cost of resolving all groups is the same, or close, to the cost of resoling one group.

But that's not the case for xCAT. There could be a number of xCAT groups that have little reason to ever be invoked via clush, and resolving all members of groups (especially dynamic groups) could be pretty costly on large systems with many groups. And in that case, resolving all groups knowing that the vast majority will never be used, doesn't seem like a very significant optimization.

There's a cluster I know where resolving all members of every single group takes much more time than cache_time :D

# time lsdef -t group -i members >/dev/null

real    3m32.317s
user    0m0.327s
sys     0m0.079s

This is largely due to the fact that resolving members of a dynamic group is an expensive SQL query in the xCAT DB (the membership is... dynamic, not defined statically) and those SQL queries are pretty much serialized, I believe.

So on a system with many dynamic groups (like this:

# lsdef -c -t group -i grouptype | awk -F: '{print $NF}' | sort | uniq -c
     31  grouptype=
    272  grouptype=dynamic

I'm not sure that a mapall call could be a real improvement, pretty much the opposite. :\

[degremont]

That's exactly the kind of data we are looking for. mapall is designed for use case where the request to get all groups is similar to get few groups (1-3). If the time is especially more expensive, no need to add it.

Do you have other scenarios, where the config is simpler? What is the perf difference in that situation?

Let's drop it. It could always be added later if that useful.

[thiell]

Thanks for the feedback! Agreed, that doesn't make sense to use mapall for xCAT in that case.