MAINFRAME > TRENDS > IBM RESEARCH

As CPU Speeds Plateau, Parallel Sysplex May Offer a Boost


In a previous article, we overviewed the pending problems with growing capacity for increasing workloads given the expected slower acceleration in single-engine speed and the difficulty of adding ever more CPUs to a z/OS image. We alluded to an approach that might be able to open another dimension of growth to bridge the gap between supply and demand for CPU capacity.

Fortunately, that approach is already available and, in fact, some installations are already using it to provide the huge capacities needed for their largest workloads. That approach is multisystem clustering and, for z/OS on System z, it’s known as Parallel Sysplex.

IBM introduced Parallel Sysplex technology at the same time as the first complementary metal-oxide semiconductor (CMOS) mainframes in 1994. It was touted then as having excellent scaling characteristics. It was said to have an overhead of about 10 percent to create a sysplex and another 0.5 percent for each additional system added to the complex. These approximations have stood the test of time and can still be used today when capacity planning for a sysplex. And, they do represent excellent scaling at very large capacities.

In the early days of Parallel Sysplex, it was viewed as a technology that provided for higher availability, easier systems management, and also as a means of achieving high levels of capacity. It’s easy to see why it was seen as a capacity enhancer in those days. The CMOS processors introduced at the same time delivered an unspectacular 15 MIPS capacity. An entire system of six processors was needed to deliver capacity approximately equivalent to a single bipolar processor of that time. Many of these little boxes would need to be clustered with Parallel Sysplex to get to a reasonable capacity level. However, this thought of Parallel Sysplex for capacity was soon overshadowed by the meteoric rise of CMOS engine speed. In just three years, a single CMOS processor delivered almost as much capacity as a whole six-processor, first-generation system. With the prospect of continued CMOS processor speed increases, systems programmers gladly avoided Parallel Sysplex implementations unless they were explicitly needed for availability. But now, as we enter an era of anemic processor speed-up, the notion of Parallel Sysplex for capacity once again becomes very interesting.

If you look at the Large System Performance Reference ratings for, say, the zEnterprise EC12 models and do a little arithmetic, you can see it doesn’t take an extremely large number of CPUs before a single-image system will deliver less effective capacity than a sysplex configuration of two systems, each with half as many CPUs. A sysplex of four systems, each with a quarter of the CPUs, can do even better. Of course, the actual crossover point will depend on workload characteristics and other factors.

A fair number of organizations already have Parallel Sysplex implementations for availability. But even these customers seem hesitant to increase the number of systems in their sysplexes, preferring instead to grow the already defined images when more capacity is needed. Some have, de facto, implemented them for capacity in that they actually have sysplexes that entail more capacity than can be delivered by a single z/OS image. Installations that have already implemented data-sharing sysplexes should, despite it being contrary to their preferences, be able to grow capacity for workloads by defining more systems to their existing sysplexes to augment whatever growth they can achieve by growing the individual member systems.

Those installations with relatively large, fast-growing workloads but without a Parallel Sysplex implementation for these workloads will face a greater challenge. They will have to deal with whatever issues have kept them from implementing the technology for the past 15 years. These include middleware that isn’t sysplex-enabled and application programs that have system affinities; these are tough issues. Fortunately, the work that’s already been accomplished in very large-capacity, single-system images will provide most installations with enough headroom to give them the time to prepare for future eventualities.

Another issue must be faced in an era when CPU speed will increase only slowly: Some applications are heavily dependent upon CPU speed. Over the decades, most major software has been updated to be able to exploit multiple processors if additional capacity is required. However, some remaining software exists that will bump into the ceiling as the amount of work needed to be done in a given amount of time continues to grow but the single-engine capacity to run it does not. Unfortunately, Parallel Sysplex technology cannot help with this sort of problem.

In summary, systems programmers have experienced a prolonged period of nearly effort-free capacity expansion, but that era appears to be coming to an end. Issues that will make it difficult for hardware and software engineers to deliver easily consumed capacity through faster CPUs or more CPUs per z/OS image might make Parallel Sysplex implementation necessary as another source of growth potential. Many of the largest installations that already have implemented data-sharing sysplexes are well positioned to exploit Parallel Sysplex capabilities. Other installations face more effort and change, but current systems and software already provide capacity levels that will give them time to prepare and make the changes to enable continued growth.

Bob Rogers worked on mainframe system software for 43 years at IBM before retiring as a Distinguished Engineer in 2012.


comments powered by Disqus

Advertisement

Advertisement

2019 Solutions Edition

A Comprehensive Online Buyer's Guide to Solutions, Services and Education.

Safely Concealed

IBM Identity Mixer is poised to change how Web users reveal personal data

Ups and Downs

IBM and Stanford University push spintronics to smaller levels

Computing in 3-D

Chips could gain depth to keep delivering on Moore’s Law

IBM Systems Magazine Subscribe Box Read Now Link Subscribe Now Link iPad App Google Play Store
Mainframe News Sign Up Today! Past News Letters