How much does a second cost?

The cost of information technology infrastructure to deliver two second response time for a given application may be much higher than the cost required to deliver three second response time.  One second of response time may have a very heavy price tag. You need to know its marginal economic cost.  If your application is performing within the performance stated in a service level agreement then it is important to understand the cost saving that can be realized in redeploying IT infrastructure while maintaining performance at acceptable levels. 

It is surprising that many service level agreements (SLAs) are drawn up without a clear understanding of their cost implications.   In such situations, quantifying the IT investment required to satisfy the prescribed SLAs would probably precipitate a change to those very SLAs. 

A non computing example clearly demonstrates the principle.  The average wait time in a telephone support queue may be required, by management, to be not more than three minutes.  If subsequent analysis determined that relaxing the requirement to a four minutes average would save 30%, and a five minute average would save 35% over the costs required to service a three minute average, then the merits of an appropriate change can be clearly presented as the cost of servicing the SLA is understood.

This page seeks to present a process for determining the savings of marginal relaxations in response time SLAs.  Note that a well behaved system, when under load, should slow down, but not stop or exhibit  a significant increase in error rates.  Reliability under anticipated load should not be negotiable (see stress testing).

In IT, the process improvement (tuning) process often stops when an arbitrary SLA is achieved, rather than taken to the next level, which may enable the redeployment of IT infrastructure.

If you have a clearly defined SLAs that covers response time of time critical transactions then you can conduct load tests to determine the minimum configuration that will meet those SLAs.  Excess infrastructure can then be redeployed.

The diagram on this page shows a three step process for determining the cost of a unit of response time.  Once the cost is known, decisions can be implemented to redeploy excess infrastructure - while maintaining the agreed response time level of service. 

Alternatively, a cost can be attributed to a given level of service, so that the cost centers of users demanding a given response time can be charged for their requested level of service. 

The First step (A) is to conduct a performance tuning cycle, by executing various types of load tests.

The improved response times that can be delivered by a tuned system (B) can then be documented.

A second series of load tests (C) can then be designed, in conjunction with system experts and architects, that determines the minimum configuration that still delivers a minimum or arbitrary level of response time performance. 

Following the process outlined in this page can facilitate the determination of cost associated with marginal changes in response time - and determining the cost or saving related to a marginal change in response time can provide a solid basis for purchase, redeployment and chargeback decisions.  

For more information on the types of tests that need to be conducted, read the Types Of Tests page and each of the links to detailed information about each test from that page.

Ask us about how we can help you design the tests that you need by emailing us at info@loadtest.com.au  and read about our consulting capacity.

Send mail to webmaster@loadtest.com.au with questions or comments about this web site.
Copyright © 2004 RPM Solutions Pty Ltd - abn 64 007 217 941             Last modified: August 04, 2004   

The content on this web page represents and demonstrates the way that RPM Solutions Pty Ltd understands and conducts load testing engagements.

This page is not an instructional guide - and RPM Solutions Pty Ltd does not take responsibility for the use of the information provided herein.