10
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. After each monitoring cycle, Enclosure Dynamic Power Capping actively
reapportions the individual power caps of the server blades based on their individual workloads. 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. It decreases the caps for server blades using less power
from cycle to cycle. Thus, the OA optimizes the power use among the server blades in the enclosure while keeping
overall power consumption below the enclosure power cap. To maintain control of the server power caps, Enclosure
Dynamic Power Capping disables external changes to the server caps using either iLO or Insight Control.
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 operates 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. Basic Power Capping cannot bring server blades under their caps within the 3 seconds
required for normal circuit breaker protection. However, the redundant side of the enclosure power can absorb the
transient overage until overall power consumption is again under the enclosure cap. If N+N power redundancy
fails, then a hardware-based failsafe mechanism overrides the Enclosure Dynamic Power Capping and immediately
lowers all server blade processors to a predetermined power state that prevents a circuit breaker overload. This
hardware override continues until power redundancy is restored.
Opting out servers
An administrator may want to leave some server blades uncapped, even though that may result in lower power
caps for the other server blades in the enclosure. Typically, uncapped servers run mission critical or consistently high
workload applications that require unconstrained power consumption. This allows them to maintain high throughput
and low latency response times. Enclosure Dynamic Power Capping allows an administrator to “opt out” of power
management for up to one quarter of the server blades in an enclosure, making them unmanaged elements. Looking
again at Figure 4, the OA measures and tracks the power these servers consume, but it is part of the power used by
the unmanaged elements pool. The enclosure power cap remains the same, so the blade power budget for the
remaining managed server blades becomes smaller.
Setting power caps for servers
Administrators can use iLO or Insight Control to set individual power caps for servers. Using Insight Control, you can
also set power caps for groups of ProLiant ML and DL servers and for groups of enclosures. For individual
enclosures, you can set power caps using the OA or Insight Control. You set power caps in exactly the same way
on servers supporting Dynamic Power Capping or basic Power Capping. Servers supporting Dynamic Power
Capping simply enforce the cap using the faster power management architecture.
