Scalability

There are a number of advantages of vertical scalability (faster processors) over horizontal scalability (more processors):

Throughput Calculations

My first college computer course was long enough ago that some of the exercises included the calculations for throughput and disk seek times. With RAID and caching nowadays, disk seek times aren't very relevant for end consumers, but the throughput calculations taught us something about CPUs that was quite counterintuitive, that is:

All else being equal, the fewer processors you have, the better. Put another way, vertical scalability is better than horizontal scalability. The reason is that vertical scalability not only gives you the same increases in throughput as horizontal scalability, but also provides increases in performance, regardless of whether or not the workload can be executed in parallel.

Example

Consider this example. Say you are sizing a system for an average load of three transactions-per-second and a maximum of five transactions-per-second at 80% utilization and that typical transactions take around 500 megacycles. You have a choice between two 1 GHz processors at $100 each or one 2 GHz processor at $200 ($100 per GHz in both cases).

If the transactions arrive at regular intervals, the response time for the two 1 GHz processors will be .5 second, and the response time for the 2 GHz processor will be .25 second.

Now consider the case where, during a given two-second interval, the six transactions arrive at two-tenths of a second intervals. What are the response times? If needed, draw three two-second long timelines, one above another, and start the transactions at the .1, .3, .5, .7, .9 and 1.1 second marks.

With two 1 GHz processors, the response time of the first two 500 megacycle transactions, which are running on separate processors, will be .5 second. The next four transactions will have a .1 second wait time and .5 second processing time, for a total of .6 second response. Therefore, the average response time will be .57 second.

With one 2 GHz processor, the six transactions will have wait times of 0, .05, .1, .15, .2 and .25 and reponse times of .25, .3, .35, .4, .45 and .5 respectively for an average response time of .37 second. So even under a varying load, the single-processor configuration is able to maintain much faster response times.

Throughput vs. Actual Bottleneck

So the moral is that, scaling horizontally by adding more processors to your system configuration or more computers to your network will only improve throughput. If the bottleneck in the system isn't throughput, you are likely to see no improvement at all.

Response Time / Page Load Time

If instead you need to improve transaction response time or web page delivery times, you need to scale the system vertically, by upgrading the system with larger, faster processors.

There's just one caveat. In systems with fairly rigid process priorities, there may be an advantage to having two processors in a single box, so that a second processor is available if one processor is tied up with a high priority task. Having a second processor available may be useful if it becomes necessary to terminate a runaway high priority task and someone needs to get through the lower priority login process because no high priority users are already logged in. So a good rule of thumb is:

Buy a single, dual-processor system with the fastest processors available, or at least fast enough to handle the expeced load. Upgrade the speed of the processors as needed and available. Consider adding additional processors or servers only if you are absolutely sure that your bottleneck is with throughput, not I/O or some other resource, and either no faster processors are currently available or else there is no advantage to improving response time or speed of web page delivery.