PD3 : Prefetching Data with DPUs for Disaggregated Memory

Description

PD3 is a memory disaggregation solution that avoids cache misses via DPU-assisted prefetching. More details can be found in our NSDI 2026 paper.

Prerequisites

PD3 requires the following dependencies:

• cmake >=3.16
• intel-tbb
• doca (version 3.2 LTS)
• ibverbs
• libaio

Hardware

Our setup involves the following hardware:

• At least 2 dedicated machines (1 for compute, 1 for memory and remote clients) : Dell PowerEdge R7625 servers with AMD EPYC CPUs (24 cores @ 2.90 GHz) and 128GB DDR5 memory, running Ubuntu 22.04
• DPU platform sits at the compute node NIC: NVIDIA BlueField-3 DPU, with a 200Gbps network interface, 16 ARMv8 A78 cores and 32GB DDR4 memory
• Network between compute / memory nodes : 200Gbps ConnectX links

Cloning

This repository uses submodules, so clone with

git clone --recursive git@github.com:fardatalab/PD3.git

Alternatively, if you have already cloned, you can run to update or fetch all submodules:

git submodule update --init --recursive

Building

This codebase uses cmake as the build system. Use the following commands to build:

On the compute node:

cd PD3/
cmake -S . -B ./build
make

On the DPU:

cd PD3/
cmake -S . -B ./build -DBUILD_DPU=ON
make

On the remote memory node:

cd PD3/
cmake -S . -B ./build
make pd3_memory_backend

Code Structure

src/
├── apps                        # executables
├── config
├── libs                        # code for experiments / other components
└── PD3                         # code for core PD3 components
    ├── buffer 
    ├── common           
    ├── literals
    ├── loading_zone            # PD3 loading zone
    ├── network_engine          # PD3 network engine
    ├── prefetcher              # PD3 host view + prefetcher
    ├── rdma_backend            # PD3 remote memory server
    ├── system         
    └── transfer_engine         # PD3 transfer engine