[dthacker]

Configuration is: p570, 4 physical CPU's
The LPAR is question is configured:
Processor Mode: Shared
Processing Units: Min=2 Desired=2 Max=3
Virtual Processors: Min=8 Desired=8 Max=12
Uncapped, Weight 180

──nmon-v10r───g=disk-Groups──────Host=omni-res───────Refresh=2 secs───14:27.36──────────────────────────────
┌─CPU-Utilisation-Small-View────────────EntitledCPU=2.00 UsedCPU 2.829─────────────────────────────────┐
│Logical                    0----------25-----------50----------75----------100                            │
│CPU User%  Sys% Wait% Idle%|           |            |           |            |                            │
│  0  12.5  13.0   0.0  74.5|UUUUUUssssss      >                              |                            │
│  1   0.0   0.0   0.0 100.0|           >                                     |                            │
│  2  25.5  10.5   0.0  64.0|UUUUUUUUUUUUsssss                  >             |                            │
│  3   5.1   0.0   0.0  94.9|UU                       >                       |                            │
│  4  16.0  12.5   0.0  71.5|UUUUUUUUssssss   >                               |                            │
│  5   0.0   5.4   0.0  94.6|ss      >                                        |                            │
│  6  27.5  11.0   0.0  61.5|UUUUUUUUUUUUUsssss>                              |                            │
│  7   0.0  11.0   0.0  89.0|sssss    >                                       |                            │
│  8  19.0  11.5   0.0  69.5|UUUUUUUUUsssss   >                               |                            │
│  9   0.0   0.0   0.0 100.0|     >                                           |                            │
│ 10  10.0  15.9   0.0  74.1|UUUUsssssss      >                               |                            │
│ 11   0.0   5.0   0.0  95.0|ss      >                                        |                            │
│ 12  41.3   8.5   0.0  50.2|UUUUUUUUUUUUUUUUUUUUssss >                       |                            │
│ 13   0.0   0.0   0.0 100.0|     >                                           |                            │
│ 14  18.5  19.0   0.0  62.5|UUUUUUUUUsssssssss>                              |                            │
│ 15   0.0   5.0   0.0  95.0|ss    >                                          |                            │
│Logical/Physical Averages  +-----------|------------|-----------|------------+                            │
│Log  11.1   8.0   0.0  80.9|UUUUUssss                                        |                            │
│Phy  47.8  38.2   0.0  14.1|UUUUUUUUUUUUUUUUUUUUUUUsssssssssssssssssss       |                            ││Entitlement Used=141.4%    +-----------|------------|-----------|------------+                            │
└──────────────────────────────────────────────────────────────────────────────────────────────────────────

1. What does the Entitled CPU measurement refer to? The desired processing units?
2. Why are physical processors 80% busy, yet logical processors are 80% idle?
3. Why do I have 16 logical processors? I thought I should have 8.
4. Could someone please explain the difference between a virtual processor and a logical processor?

I have folks in suits thinking we need more CPUs. If you run a day's worth of stats through nmon_analyzer, the CPU load never goes over 30%. I've been back through my LPAR class books trying to account for the difference between logical and physical processors, but haven't located anything.

TIA Dave

[astralvoid]

I'll try to help if I can.

1. The EntitledCPU refers to your LPAR's allocated CPUs upon LPAR startup. Generally this will be your Desired CPU amount.

2. Physical processors refer to your allocation CPU units. This refers to your UsedCPU units. Logical processors refer to the virtual CPUs that you've allocated.

3. You have 16 logical processors because SMT (Simultaneous Multi-Threading) is enabled on your requested 8 Virtual CPUs. You can enable/disable this with the smtctl -m on|off command. smtctl without args will show your current settings.

4. You can probably figure out this answer with the answers already given, but Virtual Processors are manually allocated units that you allocate during your LPAR definition. Logical CPUs will be either the same number as Virtual CPUs or 2x, depending on your SMT setting.

With regards to whether you need more CPU's, remember that you only have 4 physical CPUs. With your currently defined CPU allocations, I would imagine that you will be doing a LOT of context switching between processor time slices. My recommendation would be to lower your Virtual CPUS setting to 4 to coincide with the physical processors you have. This should reduce the amount of context switching due to there being so many Logical processors context changes during each CPU slice. You should see a better response time with this. If your application has a very heavy load of CPU intensive jobs, you may see a larger performance increase by disabling SMT (smtctl -m off) . This can be changed dynamically, so you may want to test which works best for you.


I hope this helps!! I know how complex these systems can become!

[dthacker]

Thanks for the reply. As a follow up question, is there a certain level of context switches that is considered optimal?

Dave

[astralvoid]

Based on everything I've read, there are baselines, but primarily, it is a trial and error thing.

The baseline in the documents I have state:

Uncapped Partitiions: increase # of VPU up to number of Physical processors in shared pool. (ie. 4) -- Start with 25% more than minimum and tune based on performance.

Capped PArtitions: Use minimum # VPU and monitor performance.

Bascially, each VPU is potentially worth 10ms of physical processing time. So if you allocate 1.0 CPU, and you have 1 VPU defined, the single VPU gets the full 10ms of processing time. Each additional VPU you add will split this time further, so 2VPU means each processor will get only 5 ms each. As you can see, a large number of VPUs will cause the processors to 'context switch' so much that there is very little time for real application processing.

1.0 CPU = 10ms 'real physical' processor time, thus 2.0CPU means we could have 2 VPUs with a full 10ms each since 2.0 * 10ms = 20ms or 4 VPUs with 5ms each.


I hope this makes a little sense. It's quite confusing, but once you grasp the concepts, it becomes...somewhat clearer than mud.