Abstract
A unified tiered store serves both fixed-size embedding vectors and variable-size natural-language user profiles from a shared hierarchy including GPU HBM, host DRAM, and SSD-backed persistence. A single database instance maintains separate logical partitions for embeddings and profiles with different table formats and caching representations. Cache capacity is dynamically partitioned between data types based on observed workload rates and expected value sizes with smoothing and minimum reservations. A heterogeneous I/O scheduler prioritizes embedding SSD reads and reserves SSD service for embeddings while remaining work-conserving for profile reads. Cross-type cache coherence is provided by tracking associations between user identifiers and derived embedding keys and by invalidating or version-checking cached derived embeddings when profiles update.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Convergent Memory Architecture - Unified Tiered Store for Embedding Vectors and Natural Language User Profiles", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10707