LLAMA
LLAMA is a research cache and storage subsystem that separates logical page identity from location and supports latch-free updates over a log-structured store.
Layers: A shared mapping table connects page identifiers to memory or storage. Bw-tree supplies indexed record access above this substrate; LLAMA is not itself a complete transactional database. [1] [2]
Flush protocol: Reserve buffer space before publishing a flush delta with compare-and-swap, then fill the reservation. A failed comparison leaves an invalidated reservation. Storage I/O must wait for active buffer writers. This keeps an unfinished copy from reaching storage. [1]
Durability: Flush returns after buffering, not necessarily persistence. Mk-Stable establishes durability through its supplied storage address. Internal system transactions support structural changes, not general user-transaction isolation; the access method remains responsible for WAL ordering. [1]
Recovery: A mapping-table checkpoint plus storage-log replay restores page locations. This is distinct from the transaction manager's recovery log and commit decision. [1]
Relations
Citations
- Levandoski, Lomet and Sengupta. LLAMA: A Cache/Storage Subsystem for Modern Hardware. PVLDB 6(10), 2013, pp. 877-888; sections 2-6.
- Levandoski, Lomet and Sengupta. The Bw-Tree: A B-tree for New Hardware Platforms. ICDE 2013; sections II and V.
Further reading
- Microsoft Research: LLAMA publication record, authors and August 2013 venue.
- Hesela: WAL, checkpoints and torn pages, with separate recovery guarantees.
Sources & caveats
- Primary-paper synthesis, not a native LLAMA execution or crash experiment.
- The 2013 evaluation compares Bw-tree over LLAMA with BerkeleyDB using specific workloads and hardware; no current-hardware speedup is inferred. [1]