PFA: Performance and Fairness-Aware LLC Partitioning Method

In server for cluster, the number of running programs is increasing. The benefit of consolidating multiple programs is good for server utilization, but it also leads to the program's performance degradation. Severe performance degradation can result in significant losses. Therefore, it is essential to divide the shared resource to support consolidation. Our experiment showed that even some shared resources such as CPU cores and memory had been divided, the performance of programs still drop down significantly compared to the program running alone, then we found out the primary reason was the contention for LLC. In this paper, we proposed the LLC partitioning method to improve the performance for consolidation programs. We classify the LLC usage type of the program by analyzing the LLC behavior, then allocate reasonable LLC ways according to the LLC usage type. Meanwhile, we monitor the program's performance in real-time and allocate the LLC ways dynamically. The experiment found that compared with the default LLC allocation method, our method reduced the performance loss by an average of 6.73% and improved the fairness by 0.03. Compared with the CPA method, our method reduced the performance loss by an average of 4.86%.