Read Without Fault: DPU-Mediated Memory Disaggregation over PCIe
Year: 2026
Type: conference-paper
Abstract: Data-intensive workloads can exceed local DRAM, making memory disaggregation attractive. Swap-based remote-memory systems reduce fault-handler overhead, but a demand read to a page absent from host memory still incurs a kernel fault and a full-page transfer before completing. We present Bytex, a transparent Linux remote-memory system built around a Data Processing Unit (DPU). Bytex maps evicted pages into the DPU's PCIe Base Address Register (BAR) using present, read-only page-table entries. CPU loads reach the DPU without a page fault; its programmable engines fetch remote pages into a local buffer and return the requested cache line. The demand read avoids a synchronous full-page transfer into host DRAM. Writes take a protection fault that restores a writable host copy. Device-resident pages remain immutable, allowing clean host-cache copies without additional coherence-tracking hardware. Background threads flush evicted pages and promote frequently read pages, while every host can both borrow and donate DRAM. We validate functionality with a QEMU/virtio bridge to a Hi1823E DPU and evaluate performance through simulation combining measured engine latencies with modeled PCIe costs. At 50% host DRAM relative to the working set, KV throughput reaches 1.16–1.37× Fastswap's and 1.08–1.18× Hermit's. Across the graph study, five of six GAPBS kernels reach up to 2.39× over Fastswap and 1.84× over Hermit. Trace analysis reports up to 1.56× less data movement.
Institutions: Bar-Ilan University, Huawei Technologies (China)
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225273620