MultiPath Memory Access: Breaking Host-GPU Bandwidth Bottlenecks in LLM Services

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Tang, Lingfeng, Zhang, Daoping, Chen, Junjie, Huang, Peihao, Jin, Feng, Xu, Chengguang, Chen, Yuxin, Sun, Feiqiang, Chen, Guo
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866911678363336704
author Tang, Lingfeng
Zhang, Daoping
Chen, Junjie
Huang, Peihao
Jin, Feng
Xu, Chengguang
Chen, Yuxin
Sun, Feiqiang
Chen, Guo
author_facet Tang, Lingfeng
Zhang, Daoping
Chen, Junjie
Huang, Peihao
Jin, Feng
Xu, Chengguang
Chen, Yuxin
Sun, Feiqiang
Chen, Guo
contents Host-GPU data movement has become a latency-critical bottleneck in LLM serving, surfacing in common paths such as model-weight movement and KV cache offload/fetch. Today, each host-GPU copy is effectively confined to the PCIe path of the target GPU, even though modern multi-GPU servers contain additional PCIe links on peer GPUs and high bandwidth GPU interconnects. This leaves substantial intra-server I/O capacity unused. To address this issue, we present Multipath Memory Access (MMA), a software-defined multipath memory access system for host--GPU data transfer. To the best of our knowledge, MMA is the first software-defined system to enable efficient multipath host--GPU data transfer within a single multi-GPU server. MMA expands a single host--GPU copy across available direct and relay paths without hardware, driver, or application changes. It preserves CUDA stream semantics with a dependency-preserving Dummy Task, coordinates distributed micro-transfer completion through a lightweight synchronization mechanism, and uses queue backpressure to route traffic without explicit link-state feedback. On an 8-GPU NVIDIA H20 server, MMA achieves 245 GB/s peak host-to-GPU bandwidth, a 4.62x improvement over native CUDA copies, and reduces TTFT for KV cache fetching by 1.14-2.38x and model wake-up/switching latency by 1.12-2.48x.
format Preprint
id arxiv_https___arxiv_org_abs_2512_16056
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle MultiPath Memory Access: Breaking Host-GPU Bandwidth Bottlenecks in LLM Services
Tang, Lingfeng
Zhang, Daoping
Chen, Junjie
Huang, Peihao
Jin, Feng
Xu, Chengguang
Chen, Yuxin
Sun, Feiqiang
Chen, Guo
Distributed, Parallel, and Cluster Computing
Networking and Internet Architecture
Performance
Host-GPU data movement has become a latency-critical bottleneck in LLM serving, surfacing in common paths such as model-weight movement and KV cache offload/fetch. Today, each host-GPU copy is effectively confined to the PCIe path of the target GPU, even though modern multi-GPU servers contain additional PCIe links on peer GPUs and high bandwidth GPU interconnects. This leaves substantial intra-server I/O capacity unused. To address this issue, we present Multipath Memory Access (MMA), a software-defined multipath memory access system for host--GPU data transfer. To the best of our knowledge, MMA is the first software-defined system to enable efficient multipath host--GPU data transfer within a single multi-GPU server. MMA expands a single host--GPU copy across available direct and relay paths without hardware, driver, or application changes. It preserves CUDA stream semantics with a dependency-preserving Dummy Task, coordinates distributed micro-transfer completion through a lightweight synchronization mechanism, and uses queue backpressure to route traffic without explicit link-state feedback. On an 8-GPU NVIDIA H20 server, MMA achieves 245 GB/s peak host-to-GPU bandwidth, a 4.62x improvement over native CUDA copies, and reduces TTFT for KV cache fetching by 1.14-2.38x and model wake-up/switching latency by 1.12-2.48x.
title MultiPath Memory Access: Breaking Host-GPU Bandwidth Bottlenecks in LLM Services
topic Distributed, Parallel, and Cluster Computing
Networking and Internet Architecture
Performance
url https://arxiv.org/abs/2512.16056