Description
Date depot: 3 janvier 2022
Titre: Analysis and optimization of the Linux scheduler for production environments
Directrice de thèse:
Julia LAWALL (Inria-Paris (ED-130))
Encadrant :
Jean-Pierre LOZI (Inria-Paris (ED-130))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Systèmes et réseaux
Resumé: Application software is deployed and executed in a production environment when it is user-ready and well-tested. The right choice and configuration of the production environment has a major impact on the user experience, application performance and business cost. From an in-house server room to cloud - today's businesses have a wide variety of choices for selecting their production environment. Most businesses rent their production environment from cloud service providers who present them with services like Infrastructure as a service, Platform as a service, Software as a service and Serverless. Hence the majority of today's cloud-based applications end up running in virtualized and/or containerized production environments.
Modern server hardware has also evolved quite a lot in the past few years. Dynamic voltage and frequency scaling (DVFS), Non-uniform memory access (NUMA), Non-volatile memory (NVMe) for faster I/O and Networks capable of remote direct memory access (RDMA) to name a few. But improving and optimizing the operating systems to take advantage of these new hardware capabilities still remains an important and pressing issue. According to the Linux kernel report in 2017, 90 percent of the public cloud workload runs on Linux. Currently the Linux kernel uses the completely fair scheduler (CFS), which struggles to benefit from DVFS. Many of today's popular hypervisors like Xen, KVM and Hyper-V struggle to efficiently use NUMA. Hence modern production environments impose many interesting research problems at the intersection of operating systems and virtualization and/or containerization.
CFS was traditionally designed to optimize fairness and throughput for parallel applications running on bare hardware. Hence analysis of how CFS' scheduling decisions impact the performance in virtualized and/or containerized cloud production environments makes a very interesting research area that we would like to explore during this PhD. We are particularly interested in understanding the flaws in current logic of scheduling decisions that prevent the kernel from taking advantage of new hardware capabilities, and why that is particularly harmful for production environments. Such an investigation can provide crucial help to application developers in order to select and configure the production environment for their cloud-based application. We also aspire to come up with optimizations and algorithms to eliminate those flaws and fine-tune the performance of Linux scheduler for cloud-based applications.
Doctorant.e: Pandya Himadri