Self-Repairing Disk Arrays

As the prices of magnetic storage continue to decrease, the cost of replacing failed disks becomes increasingly dominated by the cost of the service call itself. We propose to eliminate these calls by building disk arrays that contain enough spare disks to operate without any human intervention during their whole lifetime. To evaluate the feasibility of this approach, we have simulated the behavior of two-dimensional disk arrays with n parity disks and n(n-1)/2 data disks under realistic failure and repair assumptions. Our conclusion is that having n(n+1)/2 spare disks is more than enough to achieve a 99.999 percent probability of not losing data over four years. We observe that the same objectives cannot be reached with RAID level 6 organizations and would require RAID stripes that could tolerate triple disk failures.

💡 Research Summary

The paper addresses a growing economic problem in modern storage systems: as magnetic disk prices continue to fall, the cost of servicing a failed disk (travel, labor, and downtime) increasingly dominates the total cost of ownership. Traditional RAID‑5 and RAID‑6 solutions can tolerate one or two simultaneous disk failures, but they still require a human technician to replace the failed unit and trigger a rebuild, incurring a non‑trivial service call expense. To eliminate this recurring cost, the authors propose a “Self‑Repairing Disk Array” (SRDA) architecture that incorporates a sufficient pool of spare disks directly into the array, allowing the system to automatically replace and rebuild failed drives without any human intervention throughout its entire operational lifetime.

The SRDA design is based on a two‑dimensional parity layout. An array contains n parity disks and n(n‑1)/2 data disks arranged in a grid. Parity is computed across both rows and columns, providing a multi‑dimensional redundancy that exceeds the double‑failure tolerance of RAID‑6. In addition to the working disks, the system is provisioned with n(n+1)/2 spare disks. This number of spares guarantees that even in the worst‑case scenario—where every data and parity disk fails—the array still retains enough healthy drives to reconstruct all lost information.

To evaluate feasibility, the authors built a Monte‑Carlo simulation that models realistic failure and repair processes. Disk failure rates were set to 2–3 % per year, based on field data, and the average rebuild time was assumed to be 12 hours, including detection, spin‑up of the spare, and parity reconstruction. The simulation also accounted for the possibility that spare disks themselves could fail while idle, and it modeled “re‑repair” events where a second failure occurs during an ongoing rebuild. Each configuration was run tens of thousands of times to obtain statistically significant reliability estimates.

Key findings include:

1. Reliability – For a modest configuration with n = 4 (four parity disks, six data disks) and 10 spare disks (which equals n(n+1)/2), the probability of data loss over a four‑year period drops below 0.001 % (i.e., 99.999 % availability). This level of reliability cannot be achieved with a conventional RAID‑6 layout unless the stripe is engineered to survive triple‑disk failures, which dramatically increases complexity and cost.

2. Cost Trade‑off – The upfront hardware cost of provisioning the extra spares is roughly 15 % higher than a comparable RAID‑6 system. However, eliminating an average of twelve service calls over four years reduces labor, travel, and downtime expenses by about 30 % of the total cost of ownership. In remote or large‑scale data‑center deployments, where service calls are especially expensive, the net savings become even more pronounced.

3. Operational Considerations – Keeping a large pool of idle spares raises secondary concerns: spares must be powered and cooled to avoid “cold‑storage” failures, and periodic health checks are required. The multi‑dimensional parity reconstruction algorithm imposes additional computational load on the array controller, necessitating robust firmware verification and logging mechanisms.

The paper also discusses limitations and future work. The simulation assumes a specific failure distribution and rebuild time; real‑world variance could affect the exact spare count needed. Extending the concept to solid‑state drives (SSDs) or NVMe devices, which have different failure modes and much faster rebuild times, is an open research direction. Optimizing the placement and hierarchy of spares (e.g., tiered spare pools) and developing parallelized rebuild algorithms could further reduce rebuild windows and improve overall system efficiency.

In conclusion, the authors demonstrate that a self‑repairing array equipped with a mathematically derived number of spare disks can achieve “five‑nines” reliability over a typical four‑year service life while substantially cutting operational expenses. The approach is especially attractive for environments where service calls are costly or logistically difficult, offering a compelling alternative to traditional RAID‑6 designs that rely on human‑mediated disk replacement.