MetaPB: A CPU-PIM co-processing load balancer using metaheuristics algorithm.

1. Introduction

This work aims to alleviate the performance degradation issues in PIM systems caused by suboptimal scheduling. By taking into account the computational capabilities of both the CPU and PIM units, as well as the cost of data transfer between them, task scheduling is optimized from both performance and energy efficiency perspectives. This optimization enables efficient collaboration between the CPU and PIM units.

2. Prerequisites

This work is build and tested in real world commercial PIM computing system UPMEM-PIM[1] server with HW&SW environment setup below:

2.1. Hardware specifications

Category	Model	Count	Performance Metrics	Energy Consumption Metrics
CPU	Intel(R) Xeon(R) Silver 4216 (16 cores, 32 threads each)	2	2.2-3.2GHz w 22MB LLC	100W * 2
DRAM	Samsung M393A8G40AB2-CWE (64GB each)	4	3200 MHz	30W
PIM-DPU	UPMEM RISC-V DPU	2530	343-350Mhz w 64KB WRAM	100-150mW * 2530
PIM-DRAM	UPMEM BC021B (8GB each)	20	2400MHz	280W(include DPU)

2.2. Software environment specifications

Category	Name	Version
System	Ubuntu	22.04(Docker)
System	Linux kernel	5.4.0-166-generic_amd64
Toolchains	UPMEM SDK	2023.2.0
Toolchains	UPMEM driver dkms	2023.2.0
Toolchains	GCC	12.3.0
Toolchains	G++	12.3.0
Toolchains	CMake	3.28.0
Toolchains	Python3	3.10.0
Libraries	boost	1.74.0.3 ubuntu7 amd64
Libraries	xgboost	1.5.2-1 amd64

Be ware that the repository stdexec requires compiler that support C++20 standard, in order to configurate correctly, it is adviced that the CMake version is not less than 3.27.0 and with network access.

3. Overall architecture

4. Usage

4.1. Clone this repository recursively

git clone  --recursive https://github.com/A-nnonymous/MetaPB.git

4.2. Compile using cmake

cd MetaPB && mkdir build && cd build
cmake ../ && make -j

4.3. Run demos

./build/benchmraks/foo

5. References

[1] upmem