Re: per-cpu statistics (Re: CVS commit: src)

[Jason Thorpe <thorpej%shagadelic.org@localhost>]

On Apr 10, 2008, at 10:36 PM, Simon Burge wrote:

> I'm happy with keeping the API too.  Is the longer term plan to make
> _all_ evcnts (kstats/counters/whatever) per-CPU?  At collate at user
> access time?

I think there will be two classes, at least:

- Strictly per-CPU (no collating ... for stuff like "cpu0 intr count",  
etc.)

- Global with per-CPU updates, collated at user access time

And *MAYBE*:

- Global with interlocked updates.


The class should simply be declared when the evcnt is attached, and  
the API handles all of the details.  evcnts should cease being  
accessed directly by code that increments them; they should call into  
the API to manipulate the count.

I'd really prefer to not have the third class (interlocked global)  
because it just has crappy MP scalability, but might be OK for  
counters that are not modified very often.  On platforms with 64-bit  
atomic ops, it might not be TOO bad (still have the false sharing  
problem, but that's it), but other platforms would require a mutex to  
be held to update the counter.


> Simon.

-- thorpej