Users know that you can get more out of a machine by adding another processor. Most also know that by going from one CPU to two CPUs you don't get twice the power, something is lost. Typically a dual-CPU machine will have a power rating somewhere in the region of 1.90 - 1.95 times that of the single-CPU machine. What is not always appreciated is that this loss gets compounded the more CPUs you add, and a point will come when adding another processor will actually give WORSE performance than before. Where this happens depends on the magnitude of the per-processor loss, which in turn depends on the way in which the hardware manufacturer and the operating system writers have implemented the multi-processor environment. Typically the power loss per additional CPU is in the range 2-5%, sometimes even higher in older systems.

It's not just the extra raw CPU power that brings performance benefits when you add another processor. A single processor running at say, 80% utilization, will always have a number of requests for CPU resource queued up, waiting their turn. Adding a second processor will result in each CPU being at approximately 40-45% utilization, and as a result, the length of the queue at each will be dramatically shortened. The relationship between utilization and queue length is not linear - in the case above, the per-CPU queue length will almost certainly be much less than half of the value when only one processor was available.

Techniques for planning processor upgrades for computer systems vary widely. The range covers traditional techniques such as guesswork and experience, through more methodical approaches such as trending to mathematically based methods such as analytical modeling and simulation. All have their advantages and disadvantages. Only the last two will predict the non-linear effect mentioned above. Metron believe that analytical modeling offers the best compromise between cost/effort and accuracy. What do you think? E-mail your thoughts to techinfo@metron.co.uk to participate in the debate.