Make sure that metrics like CPU count support non-integer values, and ideally support concepts like cumulative CPU cycles over time, so between two time points we can determine the mean CPU clock rate. This will support virtualized partial CPUs and also variable clock rate CPUs (see AMD PowerNow) and other complicated things like Hyperthreading.
I wrote a paper for CMG06 last year called "Utilization is Virtually Useless as a Metric", I'll be presenting it at October's BayLISA meeting (venue Yahoo Santa Clara). Adrian On 9/12/07, Alexander Kolbasov <[EMAIL PROTECTED]> wrote: > We now have an OpenSolaris page for the CPU Observability project: > > http://opensolaris.org/os/project/cpufs/ > > It is currently just a small collection of notes, we should evolve it to > something real. > > I think that the most interesting question is what kind f information to > expose so that it would not encourage developers to program for non-portable > low-level details of their system and will, instead, encourage them to use the > right abstractions. > > For example, for complex CMT CPUs that have cores, strands, pipelines, etc, > should we just expose all these or should we, instead, introduce APIs that > advise the system that some threads share memory or work on different areas > of memory or some threads should be as far as possible from each other, etc. > > A big monkey wrench in the observability story is virtualization - what can we > expose that will still hold true and useful in virtualized environments? > > > - akolb > > _______________________________________________ > perf-discuss mailing list > perf-discuss@opensolaris.org > _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
