COSCUP 2024

Amidst the current zeitgeist of technological advancement, there exists a burgeoning reliance on Graphics Processing Unit (GPU)-intensive endeavors such as gaming and the training of machine learning algorithms, permeating through the fabric of myriad computer systems. While these undertakings lean heavily upon GPU prowess, their execution efficacy and performance are significantly influenced by the machinations of the Central Processing Unit (CPU) scheduler. Nevertheless, the prevailing scheduling protocols entrenched within your kernel's scheduler may prove inadequate for the exigencies of your desired workload, thereby failing to harness the full potential of your GPU or other computational resources. Enter eBPF, endowing us with the capacity to craft bespoke scheduling policies and imbue our own custom scheduler into the kernel of your operating system, sans the exigency of kernel code modification or the loading of ancillary kernel modules—processes marred by sluggishness and prohibitive costs, and often beset by compatibility conundrums. Thus, we proffer the introduction and realization of a bespoke scheduler poised to optimize performance during the rigors of machine learning model training.