Deterministic Addressing Engine (MED): Eliminating Storage Stack Overhead to Minimize Read Amplification and Energy per Useful Bit in NVMe-Based Systems
Year: 2026
Type: preprint
Abstract: Modern data-intensive workloads, including High-Frequency Trading (HFT) order books and Large Language Model (LLM) KV-caches, suffer from severe tail latency (p99) and energy inefficiency due to deep storage software stacks. Even when retrieving sub-kilobyte payloads, conventional architectures traverse B-trees, OS page caches, and 4 KB block abstractions, incurring a 64× bus read amplification for 64-byte payloads and delegating physical placement to opaque Flash Translation Layers (FTL). This paper proposes the Deterministic Addressing Engine (MED), a user-space storage architecture that replaces index-based lookups with direct mathematical offset computation (O(1)). By combining bounded-load Cuckoo Hashing (α≤0.80), kernel-bypass I/O via SPDK, and NVMe Zoned Namespaces (ZNS), MED maps keys directly to physical device offsets without internal FTL remapping. We formalize the bus amplification and energy-per-bit model (pJ/bit) and establish a two-phase validation methodology targeting order-of-magnitude reductions in p99 latency and data-movement energy.
Author: Araújo, Eriberto da Silva de
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235607123
