Is the zEnterprise is too expensive? That is a complaint DancingDinosaur frequently encounters. Acquiring a mainframe and putting into useful service is not cheap. The z114 starts at $75,000 but to run workloads you’ll have to spend more.
Usually DancingDinosaur shifts the discussion to total cost of ownership (TCO) or Fit for Purpose (or Tuned to the Task). That puts the cost discussion into the context the full costs, not just the cost of the hardware or into the context of what you’re trying to achieve. And you should be trying to achieve something; nobody buys a mainframe for the fun of it.
John Shedletsky, IBM VP of competitive technology, has been dissecting the cost of zEnterprise, Power Systems, and distributed platforms in terms of the workloads being run. It makes sense; different workloads have different requirements in terms of response or throughput or availability or security or any other number of attributes and will benefit from different machines and configurations.
Most recently, Shedletsky introduced a new workload benchmark for business analytic reports executed in a typical day, called the BI Day Benchmark. Based on Cognos workloads, it looks at the number of queries generated; characterizes them as simple, intermediate, or complex; and scores them in terms of response time, throughput, or an aggregate measure. You can use the resulting data to calculate a cost per workload.
DancingDinosaur, as a matter of policy, steers clear of proprietary benchmarks like BI Day. It is just too difficult to normalize the results across all the variables that can be fudged, making it next to impossible to come up with repeatable results.
A set of cost per workload analyses Shedletsky published back in March here avoids the pitfalls of a proprietary benchmark. In these analyses he pitted a zEnterprise with a zBX against POWER7 and Intel machines all running multi-core blades. One analysis looked at running 500 heavy workloads. The hardware and software cost for a system consisting of 56 Intel Blades (8 cores per blade) for a total of 448 cores came to $11.5 million, which worked out to $23k per workload. On the zEnterprise running 192 total cores, the total hardware/software cost was $7.4 million for a cost per workload of $15k. Click on Shedletsky’s report for all the fine print.
Another interesting workload analysis looked at running 28 front end applications. Here he compared 28 competitive App Server applications on 57 SPARC T3-1B blades with a total of 936 cores at a hardware/software cost of $11.7 million compared to a WebSphere App Server running on 28 POWER7 blades plus 2 Data Power blades in the zBX for a total of 224 cores at a hardware/software cost of $4.9 million. Per workload the zEnterprise cost 58% less. Again, click on Shedletsky’s report above for the fine print.
Not all of Shedletsky’s analyses come out in favor of the zEnterprise or even POWER7 systems. Where they do, however, he makes an interesting observation: since his analyses typically include the full cost of ownership, where z comes out ahead the difference often is not the better performance but the cost of labor. He notes that consistent structured z management practices combined to lower labor costs.
If fewer people can manage all those blades and cores from a single unified console, the z Unified Resource Manager, rather than requiring multiple people learning multiple tools to achieve a comparable level of management, it has to lower the overall cost of operations and the cost per workload. As much as someone may complain that a z114 starts at $75,000, good administrators cost that much or more.
Shedletsky’s BI Day benchmark may never catch on, but he is correct in that to understand a systems true cost you have to look at the cost per workload. That is almost sure to lead you to hybrid computing and, particularly, the zEnterprise where you can mix platforms for different workloads and manage them all in a structured, consistent way.