Answers are grounded in this paper and cite their sources.
As modern NVMe SSDs become increasingly capable of in-device computation, workloads such as SQL filtering and key-value store operations are now being offloaded from the host to the SSD. These tasks often involve small payloads, typically a few dozen to hundreds of bytes, which are inefficiently processed by the standard NVMe interface due to its fixed memory page-granularity DMA transfer. Regardless of how small the payload is, even a few dozen bytes, each request still incurs memory page-sized PCIe traffic, resulting in significant bandwidth waste and latency overhead. Existing approaches either require invasive modifications to the I/O stack and SSD internals, or offer only limited benefits for extremely small payloads. This paper presents ByteExpress, a novel mechanism that efficiently transmits small payloads by embedding them inline within NVMe queue entries, embedding for writes and extracting for reads. ByteExpress maintains PCIe traffic efficiency while delivering high performance, all while preserving full compatibility with existing APIs and SSD architectures. Our implementation and evaluation on the Linux NVMe driver and both real hardware and emulation environments, shows up to 96% reduction in PCIe traffic and 25% higher throughput compared with the standard page-based mechanism, and 79% lower latency compared with the state-of-the-art solution.
No takes yet. Share an insight, caveat, or question.
Park et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: