How Facebook’s Cold Storage Cuts Power Use by 75% While Scaling Performance

Background and Motivation

Facebook processes billions of photos daily, driving explosive growth in storage demand. Traditional cold‑storage solutions—tape libraries or optical media—are too slow, while commercial cloud options are costly. To handle petabytes of rarely accessed historical data, Facebook engineered a purpose‑built Cold Storage system.

Design Principles

Energy Efficiency : Storage nodes are powered on only when needed, eliminating redundant generators and batteries. Low‑cost commodity disks are used, and disk activity is throttled based on mean time between failures to reduce power draw.

Intelligent Management : The software must tolerate brief power interruptions without data loss, ensuring persistence and eliminating single points of failure. Metadata is self‑describing, removing the need for separate metadata services.

Future‑Proofing : The architecture anticipates scaling; performance should improve, not degrade, as capacity grows.

Hardware Architecture

The foundation is Open Vault Storage , a modular I/O topology built for the Open Rack standard. Each 2U chassis holds 30 disks (15 per tray) in a high‑density JBOD configuration, allowing interoperability with any server.

Two such systems are deployed in Prineville and Forest data centers, storing hundreds of petabytes with power consumption only one‑quarter of traditional solutions. Scaling up does not degrade performance; larger systems actually run faster.

Power‑Saving Techniques

Only one disk per tray is powered at a time, and firmware was modified to prevent accidental full‑rack power‑on events. Fan count per node dropped from six to four, power modules from three to one, and backplane lines from three to one, achieving an overall power draw of roughly 1/6 of conventional data‑center storage.

Cost‑Effective Data Protection

Instead of multiple full replicas, Facebook employs Reed‑Solomon erasure coding . Data is split into n blocks with m parity blocks; any n blocks can reconstruct the original data. Facebook’s current configuration uses a 10:4 ratio (10 data disks + 4 parity disks), providing tolerance for up to four simultaneous disk failures while using only 1.4 GB of extra space per 1 GB of data.

Facebook also runs a re‑encoding service that can adjust the data‑parity ratio as hardware reliability evolves.

Anti‑Entropy and Self‑Healing

To combat “bit rot,” a background “anti‑entropy” process scans all disks every 30 days, verifies checksums, and triggers a repair workflow that rebuilds corrupted data on fresh disks. This reduces reconstruction time from hours to minutes.

File‑System Bypass

Recognizing limitations of traditional file systems for massive, infrequent I/O, Facebook mounts disks as “bare disks,” bypassing the file system layer to gain full control over data flow and improve durability.

Scale‑Positive Performance

When capacity is added, the system rebalances data across new hardware without downtime, effectively acting like an intelligent disk‑defragmentation tool. Larger deployments thus see improved throughput rather than the usual degradation.

Future Directions

Although the two Cold Storage installations hold only about 1 % of Facebook’s total data, the approach will be expanded. Ongoing research includes integrating flash, Blu‑ray, and cross‑data‑center distributed storage techniques to further enhance persistence and efficiency.