摘要:
Methods and apparatus are provided for adaptive optimization of low-power strategies. In one novel aspect, the device monitors one or more thermal-performance parameters and determines a plurality of operation scenarios for a plurality of corresponding low-power policies. Based on corresponding operation scenarios, the device selects corresponding low-power policy. The device applies different low-power strategy for temperature control based on low-power policies. Different low-powder policy is applied to different low-power techniques, such as the DVFS, the CPU hot-plug, and the task migration. In another novel aspect, the device obtains one or more user-defined policy for each corresponding low-power technique. The selection of each low-power policy is further based on its corresponding user-defined policy. In one embodiment, the user-defined DVFS policy includes power policy, performance policy, and DVFS-balanced policy. The user-defined CPU hot-plug policy includes conservative policy, aggressive policy, and hot-plug-balanced policy. The user-defined task-migration policy includes performance policy, and task-migration-balanced policy.
摘要:
A computing system performs shared cache allocation to allocate cache resources to groups of tasks (810). The computing system monitors the bandwidth at a memory hierarchy device that is at a next level to the cache in a memory hierarchy of the computing system (820). The computing system estimates a change in dynamic power from a corresponding change in the bandwidth before and after the cache resources are allocated (830). The allocation of the cache resources are adjusted according to an allocation policy that receives inputs including the estimated change in the dynamic power and a performance indication of task execution (840).
摘要:
A computing system with multiple processor cores manages power and performance by dynamic frequency scaling. The system detects a condition when a total number of active processor cores within one or more clusters is less than a predetermined number, and an operating frequency of the active processor cores has risen to a specified highest frequency. The system also obtains ambient temperature measurement of the one or more clusters. Upon detecting the condition, the system increases the operating frequency above the specified highest frequency based on the ambient temperature measurement while maintaining a same level of supply voltage to the active processor cores.
摘要:
A computing system performs partial cache deactivation. The computing system estimates the leakage power of a cache based on operating conditions of the cache including voltage and temperature. The computing system further identifies a region of the cache as a candidate for deactivation based on cache hit counts. The computing system then adjusts the size of the region for the deactivation based on the leakage power and a bandwidth of a memory hierarchy device. The memory hierarchy device is at the next level to the cache in a memory hierarchy of the computing system.
摘要:
A switch interconnect is dynamically controlled at runtime to connect power sources to processing units in a multiprocessor system. Each power source is shareable by the processing units and each processing unit has a required voltage for processing a workload. When a system condition is detected at runtime, the switch interconnect is controlled to change a connection between at least one processing unit and a shared power source to maximize power efficiency. The shared power source is one of the power sources that supports multiple processing units having different required voltages.