Huawei Technologies Co., Limited

Liang Peng

Da Qi Ren

Ehsan Khish Ardestani Zadeh

Wei-Yu Chen

Elnaz Ebrahimi

An apparatus includes a processor core and a memory hierarchy. The memory hierarchy includes main memory and one or more caches between the main memory and the processor core. A plurality of hardware pre-fetchers are coupled to the memory hierarchy and a pre-fetch control circuit is coupled to the plurality of hardware pre-fetchers. The pre-fetch control circuit is configured to compare changes in one or more cache performance metrics over two or more sampling intervals and control operation of the plurality of hardware pre-fetchers in response to a change in one or more performance metrics between at least a first sampling interval and a second sampling interval.

Systems and methods for adaptive hybrid hardware pre-fetch

A lockstep controller operates a lockstep system of three or more CPU-GPU pairs, comparing the outputs from the CPU-GPU pairs and, by way of a majority vote, provides the output for the lockstep system. Based on comparing the outputs, if one of the CPU-GPU pairs provides outputs that disagree with the majority outputs, it can be switched out of the lockstep system. The removed CPU is replaced by a backup CPU. So that the backup CPU can be part of a CPU-GPU pair, a portion of the address space from the GPU of one of the other CPU-GPU pairs is assigned to the backup CPU to operate as a replacement CPU-GPU pair, while the CPU already associated with this GPU retains another portion of the GPU's address space to continue operating as a CPU-GPU pair.

CPU-GPU lockstep system

A fault tolerant processing environment wherein multiple processors are configured as worker nodes and redundant nodes, with a failed worker node replaced programmatically by a manager node. Each of the processing nodes may include a processor and memory associated with the processor and communicate with other processing nodes using a network. A manager node creates a message passing interface (MPI) communication group having worker nodes and redundant nodes, instructs the worker nodes to perform lockstep processing of tasks for an application, and monitors execution of the tasks. If a node fails, the manager node creates a replacement worker node from one of the redundant processing nodes and creates a new communications group. It then instructs those nodes in the new communications group to resume processing based on the application state and checkpoint backup data.

Growing community of inventors

Los Gatos, CA, United States of America

Liang Peng