On 03/29/2012 06:49 AM, Ken McDonell wrote:
I suspect it is worth making this change.
The original history (if I recall it from so long ago) was that pminfo needed to
be as lightweight as possible (the Akmal factor for those who remember) and this
was before pmprobe existed. The extra PDU round-trips to fetch and validate the
instance domains were to be avoided.
I vaguely remember some history here too - the semantics were a bit
too messy when a non-leaf is specified. e.g.
pminfo -f kernel.percpu.cpu["cpu0","cpu1"]
would be ok because all metrics in that subtree share the same
instance domain, but it gets messy when they don't. pminfo
currently allows a non-leaf to be queried and will happily
fetch all instances for all metrics in the subtree. But that
wont consistently work if a specific instance is requested.
pmdumptext just returns an error if a nonleaf node is requested:
$ pmdumptext kernel.percpu.cpu["cpu0"]
pmdumptext: Error:
r/lib/pcp/config/pmcd/pmcd.options:ib/pcp/config/pmcd/pmcd.conf: Metric name is
not a leaf in PMNS
pmdumptext: no valid metrics, exiting.
hmm, there's a bug in the error handling somewhere ...
Regards
-- Mark
|