HP ML150 HP Power Capping and Dynamic Power Capping for ProLiant servers techn - Page 11

Active power reallocation One of the more important features of Enclosure Dynamic Power Capping is the active reallocation of power amongst the server blades over time. Unlike group power capping through IPM, where a group power cap is simply apportioned as individual caps to the servers and then left alone, Enclosure Dynamic Power Capping actively reapportions the individual power caps of the servers after each monitoring cycle, based on the workloads of the individual server blades. With the blade power budget as its limit, the OA software uses a sophisticated, multi-tiered algorithm to increase the power caps of individual servers that are busier and using more power while decreasing the caps for server blades using less power from cycle to cycle. Thus the OA is able to optimize the power use among the server blades in the enclosure while maintaining overall power consumption below the enclosure power cap. To maintain its control of the power caps for the server blades, Enclosure Dynamic Power Capping disables external changes to the individual server blade power caps using either iLO or IPM. In most cases, the OA can quickly raise a low power cap for an idle server blade that receives new work. In such cases, the power sharing algorithm has little impact on performance. However, if there are too many busy server blades for the available power, the OA will attempt to share the available power fairly among all busy server blades. Enclosure Dynamic Power Capping in mixed blade environments Enclosure Dynamic Power Capping is designed to operate with all server blades that support basic Power Capping or the faster Dynamic Power Capping. It also provides circuit breaker protection using either of these types of server blades. To accomplish this, Enclosure Dynamic Power Capping relies on the extra circuit capacity of enclosures configured with N+N redundant power. Although server blades with basic Power Capping cannot be brought under their caps within the 3 seconds required for normal circuit breaker protection, the redundant side of the enclosure power can absorb the transient overage until overall power consumption is brought under the enclosure cap. If N+N power redundancy fails, then the Enclosure Dynamic Power Capping is overridden by a hardwarebased failsafe mechanism that immediately lowers all server blade processors to a predetermined lower power state that prevents a circuit breaker overload. This hardware override remains in place until power redundancy is restored. Opting out servers An administrator may want to leave some server blades uncapped, even though it may result in lower power caps for the other server blades in the enclosure. Typically, these uncapped servers are running mission critical or consistently high workload applications that require their power consumption to be unconstrained. This allows them to maintain high throughput and low latency response times. Enclosure Dynamic Power Capping allows an administrator to individually ―opt out‖ up to one quarter of the server blades from power management, making them unmanaged elements in the enclosure. Looking again at Figure 4, the power consumed by these servers will still be measured and tracked by the Onboard Administrator; however, it is now part of the power used by the unmanaged elements pool. The enclosure power cap remains the same, resulting in a reduction in the blade power budget for the remaining managed server blades. 11