Failure Recovery Issues in Large Scale , Heavily Utilized Disk Storage Systems

Large data is increasingly important to large-scale computation and data analysis. Storage systems with petabytes of disk capacity are not uncommon in high-performance computing and internet services today and are expected to grow at 40-100% per year. These sizes and rates of growth render traditional, single-failure-tolerant (RAID 5) hardware controllers increasingly inappropriate. Stronger protection codes and parallel reconstruction based on parity declustering are techniques being employed to cope with weakening data reliability in these large-scale storage systems. The first tolerates more concurrent failures without data loss at the cost of increasing redundancy overhead. The second parallelizes failure recovery from the traditional per-subsystem hardware RAID reconstruction to parallel and distributed reconstruction over all disks and RAID controllers. This paper explores the differences and similarities between large-scale storage systems in high-performance computing (HPC) and data-intensive scalable computing (DISC) for internet services, and revises reliability models for these storage systems to incorporate stronger redundant encoding and the use of parallel reconstruction. A modern example, for systems of 1-5 petabytes, suggests that triplication can have as much as 10 times lower rates of lost data per year, even when its number of components has to be almost 3 times more for the same amount of user data, but that this difference may be as small as 1 to 10 bytes lost per year. Many might decide that this factor of ten is not significant in light of other sources of information loss.