At 22:18, on Mar 14, 2006, Cornelius Bartke wrote:
Which MRTG are you using? I've found that querying SNMP for the CPU busy (.1.3.6.1.4.1.789.1.2.1.2) use will indeed graph the AVG CPU, which is the sane one.
Actually, .1.3.6.1.4.1.789.1.2.1.2 is cpuBusyTime (at least the MIB for DOT 6.5.x says so) and calculating CPU utilization via delta(cpuBusyTime)*100/delta(cpuUpTime) on a FAS960 leaves us with a no-so-satisfying ANY value. See also Bug-IDs 141441 and 148982.
It would be better to have snmpd report a cpuTable with seperate data for all CPUs.
Actually - I disagree. Graphing cpuBusyTime indeed shows the average. Been using it for years.
And that's indeed was BugID 148982 is complaining about: SNMP reports cpu utilization as the average of each CPU's utilization while sysstat reports the percentage of time at least one CPU is busy. As a result, the CPU utilization reported by the two tools may not match.
Davin.
Davin Milun 14.03.2006 23:17:
At 22:18, on Mar 14, 2006, Cornelius Bartke wrote: [...] calculating CPU utilization via
delta(cpuBusyTime)*100/delta(cpuUpTime) on a FAS960 leaves us with a no-so-satisfying ANY value. See also Bug-IDs 141441 and 148982.
It would be better to have snmpd report a cpuTable with seperate data for all CPUs.
Actually - I disagree. Graphing cpuBusyTime indeed shows the average. Been using it for years.
Hum.
We noticed a 20% increase in CPU usage when we upgraded from 6.4.4P6 to 6.5.4P1. First we were worried because there was little chance that this was due to changed usage patterns. Then we investiged by comparing sysstat output to snmpd output. Our conclusion was that in OnTAP 6.4.4P6, snmpd reports cpu utilization as the average of each CPU's utilization. With the upgrade to 6.5.4P1 snmpd reports CPU utilization as percentage of time when at least one CPU is busy, i.e. ANY instead of AVG. That has not changed when we upgraded to OnTAP 6.5.6P3.
My point is that reporting snmpd data for each individual CPU in a snmp table would leave the ANY/AVG decision to the customer and I hope you agree this would improve the usefulness of CPU utilisation data :)
-Cornelius
-
Cornelius Bartke -
Davin Milun