Exploiting Content Similarity to Improve Memory Performance in Exascale Systems.
Scott Levy, Kurt Brian Ferreira, David Fiala, Patrick G. Bridges, Dorian C. Arnold · 2012
DRAM ECC failures are one of the most frequently observed sources of node failure in large scale distributed systems. [6] As node counts continue to grow, traditional checkpoint/restart will no longer be sufficient to efficiently recover from these errors. [4] Moreover, power concerns may exacerbate this problem as we consider deploying low voltage memory chips that are more prone to error. Given that the frequency of memory errors is likely to remain a problem for the foreseeable future, we need to explore ways in which we can prevent memory errors from leading to node failures.