HiperDispatch
| Company type | Technology |
|---|---|
| Industry | Computer hardware, Mainframe computing |
| Founded | February 2008 |
| Headquarters | United States |
| Products | Workload scheduling technology for IBM mainframes |
HiperDispatch is a workload scheduling technology incorporated into modern IBM mainframe systems, such as the System z10 and IBM zEnterprise System processors, operating on recent versions of z/OS. Initially launched in February 2008, this capability was extended to z/VM starting with version V6R3 in July 2013.
A major difficulty in scaling large SMP servers lies in preserving nearly linear scalability as the number of CPUs grows. Performance improvements tend to taper off due to increased overhead from cache and memory contention, which intensifies with additional processors. With newer mainframe models supporting up to 64 CPUs (for example, the System z10), overcoming this challenge becomes increasingly important.
HiperDispatch tackles these issues by leveraging hardware enhancements, z/OS dispatching mechanisms, and the z/OS Workload Manager. Since workloads often require frequent access to memory, and in extensive SMP architectures like IBM Z, certain processors have quicker access to relevant cached data, HiperDispatch assigns work to CPUs that are more likely to hold that data in cache. When the preferred CPU is occupied, HiperDispatch opts to wait briefly instead of immediately dispatching to an idle but less optimal CPU. Nonetheless, if the Workload Manager identifies risks of missing service level objectives (such as response time goals), it intervenes by dispatching tasks to available CPUs even if their memory access is slower.
Benefit
While HiperDispatch provides marginal CPU efficiency gains on systems with a small number of processors, its impact on performance becomes substantial as CPU counts increase. IBM’s mainframe capacity planning and software licensing models assume HiperDispatch functionality is active.
Moreover, HiperDispatch "parks" logical processors so the quantity of active CPUs better corresponds to an LPAR’s assigned weight. This strategy mitigates the "short engine" phenomenon and enhances system responsiveness, including on smaller-scale setups.
Implementation
Effective operation of HiperDispatch depends heavily on correctly configuring the Workload Manager (WLM). Some environments reveal latent WLM goal conflicts once HiperDispatch is enabled; in these situations, temporarily disabling HiperDispatch may be necessary. Regardless of whether HiperDispatch is active, maintaining precise WLM policies remains vital for overall system performance.