Performance: Processor

Introduction to Processor Bottlenecks

Processor usage (also called CPU usage) is the easiest bottlenecks for performance monitor to detect. What we are looking for is the percentage time that the processor is in use rather than just ticking over by running an idle thread. If the CPU is so busy that it cannot respond to requests, then the whole server's performance soon deteriorates rapidly.

Database and email servers are the most likely to suffer from processor bottlenecks. On the other hand, file and print servers are less likely to be short of CPU power. However, large modern servers invariably have multiple symmetric processors so % Processor Time bottlenecks are becoming rarer than memory bottlenecks. That reminds me, always monitor the other major counters, Disk, Network and Memory.

If you discover a processor bottleneck, then use the process counter in performance monitor to identify which program or process is hogging the server. Also check out drivers and disk subsystems to pin point the source of the processor activity.

Note: there are two Performance Monitor counters with very similar names, processor and process, on this page we are investigating processor (CPU).
Processor Topics

* Processor: % Processor Time
* System: Processor Queue
* Multiple Processors
* Other Processor Counters
* Solutions to processor problems

As a quick way of checking processor usage, call for the Task Manger, Performance Tab. Ignore spikes but check for high continuous CPU Usage History.
Processor: % Processor Time

An overloaded processor has a distinctive and unmistakeable performance monitor profile. The % Processor Time trace looks like a curtain hanging down from an imaginary ceiling. See Diagram 1.Performance Monitor Processor bottleneck

Text books quote thresholds of between 70 - 85 percent for % Processor Time, the key point is that the counter is continuously high. It is normal for the trace to show a sharp increase when any program executes; you can safely ignore spikes.
System: Processor Queue

The hardest part of using this performance monitor counter is remembering to go to the System object (not the Processor object). What I love about any queue counters is that it is easy to remember the threshold. The rule of thumb is that the threshold for a queue bottlenecks is 2.

Other Counters

You may also wish to examine DPCs Queue/sec. This is where the server was busy so it deferred processing a request. High or intermittent bursts of Interrupts /sec could indicate a hardware problem or a loose component.

Multiple Processors

With multiple processors, it is reasonable to divide the System: Processor queue by the number of processors. So a twin processor could sustain a queue of 4.

The second rule of thumb is: you are allowed to divide the queue by the number of multiple components. (Processors, Disks or NICs)

As a point of monitoring technique, twin or quad processors give you a chance to compare Processor: %Processor Time Instances, rather than just recording the _Total.
Other Processor counters

1) Process and Thread
If you find a processor bottleneck you can pursue the cause by measuring the Process or even Thread object. What you are looking for is which instance of the Process is responsible for exhausting the processor.

2) System: %Privileged Time and Process: %Privileged Time.
The Windows 2003 operating system can execute, either in Kernel mode, which shows up as %Privileged Time, or User mode which corresponds to %User Time. This means that activities of programs like SQL or Exchange are charged to %User Time.
Here is a combination which would point to an I/O bottleneck, System: %Privileged Time > 20% and PhysicalDisk %Usage > 55%.

3) DPC
DPC means Deferred procedure calls - The processor is saying 'I am busy I will do this low priority task later.' Processor %DPC Time > 50% is suspicious and may indicate a network card bottleneck.

Solution to processor problems
Getting a second processor will work wonders for servers where the processor is being stressed. Upgrading the processor is another obvious solution for a stressed processor.
When you order the next server consider making it a quad processor. Talking of new kit, when next you spec a new system, consider clustering. Often you need two reasons to break new ground. In addition to the obvious advantage of fault tolerance, clustering can gain extra performance through load balancing across the multiple processors.