Predicting Uncorrectable Memory Errors for Proactive Replacement: An Empirical Study on Large-Scale Field Data

Uncorrectable memory errors are the leading causes of server failures in datacenters. Predicting uncorrectable errors (UEs) using the historical correctable error (CE) information helps for proactive replacement of memory hardware before the catastrophic events happen. In this paper, we perform an empirical study of UE prediction on the large-scale field data from more than 30,000 contemporary servers in Tencent datacenters over an 8-month period. We demonstrate that the traditional approach based on CE rate works poorly with a low precision. We then leverage the detail micro-level CE information to design several new predictors. The comparative study shows that the new predictor based on column fault identification boosts the baseline precision for a factor of more than 300% and at the same time also improve the baseline recall substantially.