Design and Practice of Baidu's Tape Library Storage Architecture Based on the Aries Cloud Storage System

The article introduces Baidu Intelligent Cloud's cold‑data storage architecture built on a tape‑library, originally presented at the 2023 DataFunSummit Distributed Storage Forum.

It begins with an overview of magnetic tape media and enterprise‑grade tape libraries, describing their sequential read/write nature, high reliability, low cost, mobility, and the need for specialized hardware and software.

Key characteristics of tape libraries are discussed, including massive capacity (tens to hundreds of PB), relatively low aggregate bandwidth, low idle power consumption, and the requirement to integrate with vendor‑specific management software.

The paper then outlines typical tape‑library use cases such as backup, archival, cold‑data storage, and innovative applications like large‑capacity recycle bins.

Next, the Aries cloud storage system (A Reliable and Integrated Exabytes Storage) is introduced. Aries provides three data models—Slice, Volume, and Stream—and is composed of four subsystems: resource management, user access, tape‑library storage, and verification/repair. Aries manages petabytes of data across thousands of servers.

Aries' data and access models are explained: Slices are the smallest immutable units, Volumes aggregate slices without order, and Streams preserve slice order. Access patterns for each model are described.

The tape‑library architecture is detailed, emphasizing four design principles: physical‑aggregate writes, decoupling of user writes from tape‑library dumping, location‑aware retrieval scheduling, and reuse of existing tape‑library software capabilities.

A data‑flow diagram shows how user data first lands in Aries' disk pool, is asynchronously dumped to tape, and later retrieved back to the disk pool before being served to the user.

The aggregation‑write process is explained, including the allocation, writing, sealing, and size‑checking of Volumes (8‑16 GB) that are later dumped as linear files onto tape.

The dump process consists of reading sealed EC Volumes, appending their slices to a linear file on a GPFS/StorNext filesystem, and invoking LTFS‑EE to migrate the file to tape with two‑copy redundancy.

The retrieval process involves task persistence, location‑based scheduling, LTFS‑EE recall to a local cache, slice extraction, and writing the slices back to the original EC Volume in the disk pool.

A real‑world case study describes a business with petabyte‑scale cold data requiring long‑term, low‑cost storage. The solution deployed two 102 PB LTO‑8 tape libraries (44 drives each) organized into logical pools, with GPFS providing a unified namespace.

Hardware details of the tape‑library head servers are listed (dual HBA cards, dual NICs, Optane RAID‑1, etc.), and the software stack (LTFS‑EE, GPFS, TapeNode) is outlined.

Performance results show steady write rates of 0.8‑0.9 PB per day until the library filled in early 2023, and retrieval benchmarks where 124 volumes (~1 TB total) were restored in an average of 14 minutes, demonstrating that tape‑library retrieval can be much faster than traditional expectations.

The article concludes with a summary of the architecture, design rationale, and operational outcomes.