Fairness in Reliability: Equitable Use of Reliability Enhancement in Caches

Abstract

Continuous technology scaling has made modern computer systems increasingly vulnerable to soft errors, making fault tolerance approaches increasingly important. Although such approaches generally provide substantial reliability enhancements, their performance and energy overheads can be prohibitive. Selective fault tolerance methods address this problem by providing sufficient reliability instead of conservative (worst case) reliability. However, when multiple applications use the same architecture with limited reliability enhancement capability, allocating this capability across competing applications in a fair manner remains an open research problem. Our paper addresses fairness problem in the reliability domain by focusing on cache memory. The proposed framework dynamically allocates ECC-protected cache blocks to concurrently running applications based on their error rates. To achieve this, errors are monitored at three levels-physical, data access, and application-each employing a distinct reliability metric. The effectiveness of the framework is evaluated in various scenarios as part of our empirical study. Our framework successfully enhances fairness in reliability, particularly when applications exhibit comparable error resilience. In heterogeneous workloads, fairness in reliability ($FiR$), calculated as the error rate difference across applications, is improved from 0.48 to 0.29; on the other hand, for homogeneous workloads, the $FiR$ value is improved from 0.67 to 0.30. Experimental results indicate that error rates obtained from lower levels allow us to adjust the error rates at the application level, resulting in fairer usage of protected hardware resources.