Stephen Hemminger [off-list ref] writes:

This is crap. only a small fraction of these SNMP counters are
close enough to the hot path to deserve per-cpu treatment.

Interesting.  I was just reading af_inet.c:ipv4_mib_init_net()
to get a feel for what a large consumer of percpu counters looks
like.

I expect the patterns I saw will hold even if the base size is much
smaller.


Your statement reinforces my point that our allocations in a per cpu
area are small and our dynamic per cpu allocator is not really
optimized for small allocations.

In most places what we want are 1-5 counters, which is max 40 bytes.
And our minimum allocation is a full cache line (64-128 bytes) per
allocation.



I'm wondering if dynamic per cpu allocations should work more like
slab.  Have a set of percpu base pointers for an area, and return an
area + offset in some convenient form.

Ideally we would just have one area (allowing us to keep the base
pointer in a register), but I'm not convinced that doesn't fall down
in the face of dynamic allocation.

Eric