Home >> White Papers >> Balancing Server CPU & I/O w/ FlashDisk RAID

Balancing Server CPU & I/O w/ FlashDisk RAID

Background
Computers used as servers today have 100 times more processing power than they did just a few years ago. Now too, it is possible to add multiple processors to again multiply processor performance to unparalleled levels. However, mechanical disk drives have improved only marginally, perhaps tripling in speed in the same time period. Thus servers today, especially file and print servers, are often 90+% idle because they are starved for data to process. A few years ago, a typical application, whether batch processing or database inquiry, was often well balanced in its demands on the system. The chart shows that a typical "1995" job spent 50% of its time in processing and 50% of its time accessing data. Not so anymore. The "2000" job in the chart is typically spending 90% of its time in I/O processing.

Disk I/O Bottleneck
Consider the impact of today's new processor running 100 times faster. This new processor slashes the time spent on processing the job from 50% to just 0.5% of the original benchmark time-the direct result of 100 times greater CPU performance. At the same time, today's disk drives cut I/O time by a factor of three, from 50% to 16.7%. The resulting job execution time is now reduced to 17.2% of the original 100%. This is pretty impressive until you consider that today' jobs and user workloads have expanded with more sophisticated databases, client/server architectures, high resolution graphics and other advanced features to fully consume this improvement and more. A closer examination of the remaining workload distribution shows that 0.5% of the time is spent in processing and 16.7% of the time remains in disk I/O-thus 16.7/17.2 or 97% of the job in this example will be I/O dependent. This is a classic disk I/O bottleneck that is caused by the remarkable improvement in CPU speed-and corresponding lack of comparable improvement in disk I/O speed.

Diminishing Returns
If somehow, the processor were to increase in speed ten-fold again to 1,000 times the 1995 speed, the job would only improve from 17.2% of the old benchmark to 16.75% -- a tiny improvement from a major investment -- and a classic case of diminishing returns. The resulting job would then spend 99.7% of its time performing disk I/O. Clearly, the time has come to address the "other half" of the work that now accounts for typically 80% to 99% of the total execution time.

Invest in CPU or I/O?
Since a typical server job is now often 90% or more of the remaining work, any improvement in the disk I/O translates nearly directly into total overall speed up of the entire application. For example, suppose that you have a choice between a new CPU that is five times faster and a new disk I/O system that is five times faster? Which is better? Many network managers now automatically adopt every new CPU improvement from habit and because doing so has worked well for them in the past-not to mention the influence of sales and marketing from server manufacturers. The situation is different today. A new CPU that runs five times faster will cut the 10% portion of the job to 2%-or just an 8% overall savings in job time. A new disk I/O system that runs five times faster cuts the 90% portion to just 18%-a 72% savings. The job with the I/O improvement now runs in 72% less time or over 3.5 times faster than just prior to the investment and 3.3 times faster than the job with the CPU investment.

Return on Investment
Typically a five-fold increase in processing power is available only at significant cost. Suppose on a mid-range server with 200 users, this costs $20,000 extra, or $100 per user. Each user saves 8% in job processing time. Now suppose that the fast I/O system also costs $20,000 but that the storage is needed anyway. Assume that standard storage could be purchased less expensively for $15,000 and that $5,000 extra was spent to obtain the speed. Thus the speed was purchased for only $25 per user. Each user saves 72% in processing time.

The comparative analysis is now obvious. A 100 minute job will take 92 minutes after the CPU investment and 28 minutes after the I/O investment. Thus the 28 minute job will run about 3.3 times as

fast as the 92 minute job. For one quarter of the investment per user, the performance improvement is 3.3 times as large which is 13.2 times (1,320%) better return on investment. The better investment choice is clear -- add disk I/O capability for the best return on investment. Further, by investing in disk I/O performance, you rebalance the system and you are unlock the CPU investment that you already paid for.

Payroll Savings
The results are easily translated into direct payroll savings as well. Suppose that the users run these and similar applications including database updates and transaction processing tasks all day long and yet, conservatively, average only 25% of the day actively on the computer. A saving of 72% of the application time for only 25% of the day still yields an overall 18% time savings. If the average person earns $36,000 per year in payroll and benefits, this translates into $6,480 annual savings per user -- on a mere $25 per user investment. This investment pays

back in just a few days. The CFO will be ecstatic. Management can postpone or reduce new hiring, reassign workers to new projects, work off the backlog of pending projects or reduce work force by attrition and pocket the savings that are real and tangible .

More Valuable Benefits
Many other benefits are also likely to accrue. A company may produce a product earlier and thus capture higher market share, higher revenues and higher profits. Extra projects tackled with the extra time available to individual users could open up entire new revenue streams or cost reductions. The benefits are limited only by the enterprise's ability to use the extra resources.

contact us | site map | privacy
copyright ©2008 by Winchester Systems