| Title: | POLYCENTER Performance Data Collector & Advisor for OpenVMS |
| Notice: | Latest version V2.2-51 |
| Moderator: | BSS::JILSON |
| Created: | Wed Nov 07 1990 |
| Last Modified: | Thu Jun 05 1997 |
| Last Successful Update: | Fri Jun 06 1997 |
| Number of topics: | 2413 |
| Total number of notes: | 9017 |
Hello,
customer has some difficulty understanding the CPU UTILIZATION with
respect to the TD_SINGLE_CPU_VUP value used in the Augmenting Knowledge
base file.
Consider the following:
FIRST.VPR
THRESHOLD
TD_SOFT_FAULT_SCALING_419 = 15.00
TD_IMGACT_SCALING_419 = 7.5
TD_HARD_FAULT_SCALING_419 = 1.7
TD_COM_SCALING_419 = 2.10
TD_SINGLE_CPU_VUP_419 = 10.0
ENDTHRESHOLD
$ advise performance compile FIRST.VPR
$ advise performance graph /type=cpu_utilization=(stack)/rules=FIRST
SECOND.VPR
THRESHOLD
TD_SOFT_FAULT_SCALING_419 = 15.00
TD_IMGACT_SCALING_419 = 7.5
TD_HARD_FAULT_SCALING_419 = 1.7
TD_COM_SCALING_419 = 2.10
TD_SINGLE_CPU_VUP_419 = 50.0
ENDTHRESHOLD
$ advise performance compile SECOND.VPR
$ advise performance graph /type=cpu_utilization=(stack)/rules=SECOND
His question:
shouldn't the SECOND CPU graph show 5 times LESS the values compared
to the FIRST CPU graph ?
(using the same .CPD file for creating the graphs)
I did some testing and found no difference between graph FIRST and SECOND.
Are there any 'hard-coded' values used within the used files and/or has
the Data Collector any 'knowledge' about these values?
regards,
Rob Boor, Off-Site Services, Utrecht - The Netherlands.
| T.R | Title | User | Personal Name | Date | Lines |
|---|---|---|---|---|---|
| 2396.1 | BSS::JILSON | WFH in the Chemung River Valley | Tue Feb 25 1997 15:33 | 5 | |
This should only effect this graph if there are multiple nodes being graphed at the same time. Then the VUP rate would be used to normalize the graph. From .0 I don't take that there are multiple nodes being graphed. Jilly | |||||
| 2396.2 | Still questions | UTRTSC::BOOR | Wed Feb 26 1997 07:45 | 40 | |
Jilly,
thank you for the quick explanation, but still having some questions.
Customer upgraded their VAX7640 cluster to a VAX7840 cluster, running
OpenVMS V6.2. These VAX78nn's are not fully supported with V6.2 and
are reporting itself as HW_MODEL 419 (VAX7740) instead of the correct
HW_MODEL 502.
The full support is within V7.1 of OpenVMS, but this customer can't upgrade
to this version because of production environment (Amsterdam Exchanges !)
We want to get correct graphs to predict any performance bottlenecks and
take appropiate action. (mostly buy HW...)
We thought we could get the correct graphs by using the 7840 numbers
(got from CA, by raising IPMT level 2 case) and put them into a threshold
file using HW_MODEL 419
THRESHOLD
! Hardware model number 499 values received from CA
TD_SOFT_FAULT_SCALING_419 = 12
TD_HARD_FAULT_SCALING_419 = 1.7
TD_IMGACT_SCALING_419 = 6
TD_COM_SCALING_419 = 1.8
TD_SINGLE_CPU_VUP_419 = 69.2
ENDTHRESHOLD
! The same needs to be done for all models from 499 to 514.
During the 7640 period customer had an average CPU utilization around 60%,
(7840 is 200% faster) so they were now expecting a CPU utilization around
30%.
But the graphs are showing an average of 48%, this is the 7740 scaling
value.
(DECPS values: VAX7610 - 34.0 / VAX7710 - 40.0 / VAX7810 - 69.2)
Can we make a work-around/hack for this customer to make him happy again.
He was very disappointed by the fact it took FOUR weeks to get the
78nn values from CA by raising a IPMT level 2 case.
Rob.
| |||||
| 2396.3 | BSS::JILSON | WFH in the Chemung River Valley | Wed Feb 26 1997 11:57 | 20 | |
For a single system cpu utilization is simple to calculate. for a given time period there are T number of seconds of cpu time available per cpu and PSDC collects the actual number of seconds consumed on each cpu. Therefore VUP rating is completely irrelvant. The amount of time consumed is a known fact and therefore cpu utilization is an exact calculations - total number of seconds of cpu time consumed -------------------------------------------- time period being measured X # of cpus In your case the amount of cpu time consumed didn't scale exactly with the change in relative cpu power. This can happen for many reasons. The biggest is that the customer has added more load but just doesn't realize it. The second biggest reason is that the change in cpus has removed a bottleneck and is reflected in the larger amount of cpu time consumed than what one would expect. Without a complete performance analysis of the before and after time periods one cannot say exactly what is causing the larger than expected cpu consumption. Jilly | |||||