Missing Data Recovery in Large-Scale, Sparse Datacenter Traces: An Alibaba Case Study

摘要: The trace analysis for datacenter holds a prominent importance the performance optimization. However, due to error and low execution priority of collection tasks, modern traces suffer from serious data missing problem. Previous works handle recovery via statistical imputation methods. such methods either recover with fixed values or require users decide relationship model among attributes, which are not feasible accurate when dealing two trends in traces: sparsity complex correlations attributes. To this end, we focus on released by Alibaba propose tensor-based facilitate efficient large-scale, sparse traces. proposed consists main phases. First, discretization attribute selection work together select attributes strong value-missing attribute. Then, tensor is constructed recovered employing CANDECOMP/PARAFAC decomposition-based completion method. experimental results demonstrate that our achieves higher accuracy than six machine learning-based