Just a short historical note ... PM_CONTEXT_LOCAL was invented in early
1994 to address a very specific need, namely Akmal Khan (my boss at the
time) would not buy this whole pcp story unless we could collect
performance data in something close to the overhead and costs of
vmstat(1).
PM_CONTEXT_LOCAL avoided a couple of context switches and some pipe
traffic in the steady state and that was why it was invented ... it is a
little surprising to see this ugly hack still being used 15 years on!
The other remaining early and misguided optimization that survives is
sample.sysinfo which was another attempt to achieve the same goal.
I wonder if anyone other than the p76 pmda ... which is no longer
required ... 8^(> ... is using the other pcp dinosaur in PDU_ASCII?
On Wed, 2009-03-11 at 13:06 -0500, Martin Hicks wrote:
> On Wed, Mar 11, 2009 at 11:06:58AM +1100, Nathan Scott wrote:
> > ----- "Martin Hicks" <mort@xxxxxxx> wrote:
> > > Hi,
> > >
> > > 2nd crack a fixing the IB memory leak.
> > > Also I think I fixed the problem with the clusterpmda and the IB
> > > metrics.
> > >
> > > I think this really highlights the fact that every PMDA that supports
> > > being loaded via a local context should use unique names for its
> > > variables, such as 'metrictab'. I'll send in a patch soon that
> > > addresses this problem with the ib PMDA.
> >
> > Does making metrictab "static" fix the problem? That seems simpler,
> > if so...
>
> I don't think so. We autogenerate a metric_table.c file by walking
> through the metrictabs from the linux and IB pmda, so the metrictab is
> in a separate file.
>
> mh
>
> _______________________________________________
> pcp mailing list
> pcp@xxxxxxxxxxx
> http://oss.sgi.com/mailman/listinfo/pcp
|