A Method for Acquiring and Analyzing Program Resource Sensitivity for Workload Consolidation

A method for acquiring and analyzing program resource sensitivity is proposed to deal with the problem that the performances of programs and systems are unstable due to workload consolidation in clusters or data centers. A control group with tunable resources is set up by using Cgroups in Linux, and programs are executed in the control group under various resource conditions. The performance of programs within limited resources is obtained and used for calculating the resource sensitivity. The sensitivity values of programs under restriction of multiple resources form a sensitivity hyperplane to predict the performance of co-running programs. The resources required by a program are reordered according to their importance for cluster scheduling. Experimental results show that the average absolute error between the real sensitivity value and the sensitivity value predicted by the method is 0.093. A comparison with the common least load scheduling strategy shows that the scheduling strategy using resource sensitivity aware improves program performance by 22.2% and enhances both disk I/O and network I/O utilization. It also shows the contribution of program resource sensitivity to workload consolidation.