Allocation of Partitions to Nodes in a Power-Aware Computing System
Publication Date: 2010-Nov-15
The IP.com Prior Art Database
A method and system for minimizing the number of nodes powered on by allocating partitions to nodes in a multi-node system based on power, frequency, and performance criteria.
Page 01 of 4
Allocation of Partitions to Nodes in a Power -Aware Computing System
Disclosed is a method and system for minimizing the number of nodes powered on by allocating partitions to nodes in a multi-node system based on power, frequency, and performance criteria.
Various packing schemes to reduce power and save energy have been implemented for single-node systems. Extending these schemes to multi-node systems pose a set of unique problems for system-level power management since each node has its own power management controller, and there is no direct coordination between the controllers for each node.
Power usage can be significantly reduced by minimizing the number of nodes that a machine has to have powered on. However, the allocation of nodes must try to avoid violating any pre-existing power caps on the nodes that it uses unless the caps are soft caps and can be increased. A combination of the power management system and power-aware partition management has to ensure that partitions contain processors which all run at approximately the same frequency. Finally, nodes vary in their power and performance efficiency due to component-level variation and other design considerations. This creates an opportunity to use the most efficient nodes first and use the less efficient ones only when the machine is more fully populated with partitions.
Considering these factors provides the motivation for this method and system for allocating partitions to nodes in a multi-node system based on their defined sizes. It assumes that the system is capable of power shifting across the nodes. It does not consider the case of adhering to fixed power cap assignments to nodes. It re-assigns the partitions to nodes when it encounters a change in the number of active partitions, an alteration of the definition of any active partition, or a hard node power cap. Such hard node power caps are generally either the result of machine design choices or a system-level power cap. The primary advantage of this method is that it reduces the overall system power and energy and minimizes the constraints imposed by the presence of cross-nodal partitions.
Partitions are allocated on a multi-node, power-managed computer system in a power-aware manner with goals in the following in order of decreasing importance:
Minimize the number of nodes that the system has to have powered on. Thus, the allocation only uses nodes when the nodes already powered on do not have the resources to accept the next partition definition.
Minimize and, if possible when consistent with the first goal, eliminate cross-nodal partitions. Each node is assumed to have its own power-management controller and that there is no direct communication path or coordination between power-management controllers. As a result, since all partitions must maintain no greater than a 10% difference in processor frequency between the fastest and the slowest processor core being used by the partition, cross-nodal partitions...