On Tue, 18 Mar 2008 06:13:11 -0700
Hal Rosenstock <[EMAIL PROTECTED]> wrote:
> Eli,
>
> On Tue, 2008-03-18 at 15:06 +0200, Eli Cohen wrote:
> > Hal, can you explain the mechanism that can be used to get counter
> > values larger than 32 bits? I see that the IB spec defines the counters
> > as 32 bits.
>
> Polling (at a configurable rate), clearing above some threshold, and
> aggregation into larger counters on the manager side. For OpenSM, see
> perf-manager-arch.txt or the current implementation: osm_perfmgr.c and
> osm_perfmgr_db.c
>
./configure --help
...
--enable-perf-mgr Enable the performance manager (default no)
...
./opensm --help
...
--perfmgr
Start with PerfMgr enabled.
--perfmgr_sweep_time_s <sec.>
PerfMgr sweep interval in seconds.
...
Run opensm with the console for more control of the perfmgr and to dump the
counters.
OpenSM $ help perfmgr
perfmgr [enable|disable|clear_counters|dump_counters|sweep_time[seconds]]
perfmgr -- print the performance manager state
[enable|disable] -- change the perfmgr state
[sweep_time] -- change the perfmgr sweep time (requires [seconds] option)
[clear_counters] -- clear the counters stored
[dump_counters [mach]] -- dump the counters (optionally in [mach]ine
readable format)
[print_counters <nodename|nodeguid>] -- print the counters for the specified
node
OpenSM $ perfmgr
Performance Manager status:
state : Enabled
sweep state : Sleeping
sweep time : 180s
outstanding queries/max : 0/500
loaded event plugin : opensmskummeeplugin
Dump counters dumps to the file configured in opensm.opts.
...
#
# Event DB Options
#
# Dump file to dump the events to
event_db_dump_file /var/log/opensm_port_counters.log
...
There is no guarantee against losing counts, however, if you set the sweep time
fast enough it should be pretty accurate.
Hope this helps,
Ira
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
