CONCEPT Cited by 1 source

Tiered cache as planetary memory

Definition

Tiered cache as planetary memory is the architectural analogy that models a globally distributed storage system as a CPU memory hierarchy at planet scale — with local memory/flash as L1/L2, regional storage as L3, and global storage as "disk" — with transparent on-demand hydration between layers, just as an OS kernel hydrates data through page cache and CPU caches.

The Analogy (Meta's Framing)

What if we think of storage as a disk in a planet-scale computer and borrow ideas from the operating-system world?

• L1 cache: GPU host memory (distributed peer cache via Owl)
• L2 cache: GPU host flash (local SSD)
• L3 cache: Regional BLOB-storage fabric (flash-backed)
• Disk / Source of truth: Global BLOB-storage fabric (HDD-backed)

When a dataloader reads, data is hydrated on demand through these layers transparently — just as a Linux process reading from disk triggers transparent hydration through page cache → L2 → L1 CPU caches.

Enabling Mechanisms

1. Dataloader prefetch: next batch prefetched into L1 during current batch processing
2. Deep prefetch API: explicit prefetch() triggers hydration from global → L3 + metadata cache warmup
3. Automatic data lifecycle: TTL and LRU eviction policies, capacity/quota-aware

Impact

Eliminated hours of explicit data ingestion time. Researchers ingest data once to global storage and access it from any region without thinking about regional boundaries.

(Source: sources/2026-07-01-meta-ai-storage-blueprint-at-scale, "Maximizing Research Velocity" section)

Seen in

• sources/2026-07-01-meta-ai-storage-blueprint-at-scale — canonical instance at Meta