Stella Bitchebe

Research Abstract:

PhD Dissertation: Out-of-Hypervisor Virtualization. I acquired a strong Systems background during my Ph.D., where I specialized in virtualization. I proposed a new research axis called Out of Hypervisor (OoH). In virtualized clouds, OoH exposes hardware virtualization features to guest Operating Systems (OSes). I illustrated OoH in my work with two recent virtualization features from Intel: PML (Page Modification Logging) and SPP (Sub-Page write Permission), and four use cases: working set size (WSS) estimation of virtual machines (VMs), checkpoint/restore (CR), garbage collection (GC), and memory vulnerability prevention applied to buffer overflow mitigation in guest userspace. For WSS estimation, I introduced PRL (Page Reference Logging), an extended version of PML that provides efficient and accurate estimation with no performance degradation on users’ applications. PRL is published at VEE 2021, and its artifacts are available on GitHub. For CR and GC, I used OoH principles to expose PML (OoH for PML) for accelerating process checkpointing (e.g., CRIU) and concurrent garbage collection (e.g., Boehm GC) in the guest. OoH for PML has been published at SC 2022, and its source code and artifact are also available on GitHub. For memory vulnerability mitigation, I introduced GuaNary, a novel safety guard against overflows, allowing synchronous detection at a low memory footprint cost compared to the state-of-the-art. GuaNary leverages Intel SPP and employs OoH principles to expose it to guest processes. It is under revision (1-shot revision)at SIGMETRICS, and the source code is available on GitHub. OoH results have proven its concrete benefit for cloud users. For example, the SC paper shows that applying OoH for PML can bring, on average, up to 4× speedup to HPC applications compared to the prior CR techniques. This with an average overhead of only 3.5%, an acceptable threshold for production environments. Similarly, applying Guanary to state-of-the-art memory allocators can allow for protecting 25× more buffers in cloud applications, with 60% less memory consumption and only 7.7% overhead on average, which is also acceptable. Current Research Path: OS specialized for ML. The computer kernel, context switching, memory management, etc., are just some of the OS/kernel components that can play an essential role in the performance of data preprocessing for Machine Learning (ML) workloads. My current research project explores the path to a specialized OS for ML with special attention to memory management for ML’s data preprocessing. For example, irregular memory access patterns can be a limiting factor to the overall performance of the hardware accelerators, and if many applications run simultaneously on the OS, the swapping mechanism induced by the LRU (Least Recently Used) list management of the Linux kernel can affect the preprocessing speed.

Bio:

I am a Postdoctoral Researcher in the School of Computer Science at McGill University, working with Prof Oana Balmau in the Data-Intensive Storage and Computer Systems Laboratory (DISCS Lab). I also hold the status of invited researcher at the LIG laboratory at Université Grenoble Alpes. My main research axes are Systems and Machine Learning (ML). With a background in Systems and a recent interest in ML, I aim to work at their intersection, proposing systems-level optimizations for ML frameworks and algorithms. I received my Ph.D. in France at Université Côte d'Azur under the supervision of Professor Alain Tchana. My Ph.D. research mainly focused on Operating Systems and Hardware Virtualization. I proposed a novel virtualization research axis called "OoH: Out of Hypervisor." Instead of emulating full virtual hardware inside a virtual machine (VM) to support a hypervisor, the OoH principle is to individually expose current hypervisor-oriented hardware virtualization features to the guest OS. I received my Diploma degree in Computer Science at the Polytechnic School of Yaounde in Cameroon. I am involved in promoting STEM in general, particularly computer science, to young girls (both high school and university students). I have set up an association to tackle academic harassment in Cameroon. Last year, I launched an initiative in Cameroon named WoCC (Women in Computer Science Cameroon), an annual seminar seeking to expose aspiring women scientists to women’s accomplishments worldwide in computer science research.