NeoMem-MICRO-2024

The Artifact of NeoMem: Hardware/Software Co-Design for CXL-Native Memory Tiering. This project is developed on Intel's CXL FPGA platform

Overview

1. Introduction

2. Hardware Requirements

3. Installation

4. Run Experiments

5. Check Results

1. Introduction

NeoMem project consists of three submodules: NeoMem_FPGA, linux and experiments. NeoMem_FPGA contains the FPGA code of NeoProf. The linux folder contains NeoMem's host-side driver, daemon and user-space interface implementation. You can find some click-to-run scripts in the NeoMem_benchmarking folder, which helps you reproduce the main results of NeoMem paper.

2. Hardware and Software Requirements

Here are the hardware and software setups used in NeoMem:

1. Supermicro SYS-741GE-TNRT platform
2. Intel Agilex-7 I series FPGA Dev Kit @400Mhz
3. 32GB DDR5 4800Mhz DRAM x4
4. Single Socket Xeon 4th 6430 CPU @ 2.1GHz
5. Quartus 22.3 with CXL 1.1 IP

3. Installation

To download this repo, run:

git clone [email protected]:PKUZHOU/NeoMem-MICRO-2024.git
cd NeoMem-MICRO-2024
git submodule update --init

To install the provided linux kernel, run:

cd linux
make -j
make modules_install
make install

Note that we hardcode some address in our driver (in linux/drivers/neoprof/neoprof.c), you may need to change it on your platform.

NEOPROFILER_MMIO_BASE defines the start address of the BAR space of the CXL-enabled FPGA device. To find the address on your platform, run:

lspci -vvv

You can find the address as follows.

CXL_MEM_BASE defines the start memory address of the CXL-enabled FPGA device. To find the address on your platform, run:

cat /proc/iomem

The CXL memory device is detected as a DAX device. You can find the address as follows.

4. Run Experiments

To get the result shown in Fig.11 and Fig.13, run

cd NeoMem_benchmarking
python run_all.py

The experiment result will be stored in output/experiment_output

To get the result shown in Fig.16, run

python run_convergence_curve.py

The experiment result will be stored in output/gups_convergence_analysis

5. Check Results

To visualize the experiment result, run

cd NeoMem_benchmarking
python parse_result.py

The result figure will be stored in output/fig_output