Adaptive Coarse Grained Cloud Computing Management Mechanism for Collaborative Filtering

摘要

Under cloud computing of multimedia application environment, the nuclear calculation failure has great impact on the normal operation of program. To alleviate the impact of failure, a roll-forward calibration method with fault-tolerant cloud computing and core-failure calibration of parallel data package is put forward. The proposed scheme includes a new integer matrix multiplication and convolution operation. The specific way is that the input data is provided with symmetry package, and corresponding GEMM operation is conducted by using package data. Due to packing output containing redundant information, the available output can alleviate a certain number of continuous kernel failures. What matters is that this method does not need additional calibration input processing. Therefore, compared with existing FER method, core-failure recovery mechanism is first put forward, without additional processor core (for example, for calibration or modular redundant calculation).