ICSA Colloquium Talk - 08/06/2021 Title: Fast Local Page Tables for NUMA Servers with Mitosis Abstract: Increasing memory heterogeneity mandates careful data placement to hide the non-uniform memory access (NUMA) effects on applications. While NUMA optimizations have focused on application data for decades, they have ignored the placement of kernel data structures due to their small memory footprint. Yet, misplacement of kernel data structures in large NUMA systems can lead to significant slowdown of application performance. In this talk, I am going to demonstrate the adverse effect of misplaced kernel data structures at the example of page tables. To quantify the possible slow down, we conduct a systematic study of various allocation policies for both, the data pages and page-table pages, and its effects on the application's performance. Moreover, we show that in a multi-threaded application, memory accesses to remote NUMA nodes in during to page-table walks are inevitable. Based on our findings, we present Mitosis a set of kernel and hypervisor extensions that eliminate remote memory accesses during page-table walks. Mitosis replicates and/or migrates the application's page table across the NUMA nodes of the system. With Mitosis, applications observe a significant speedup due to improved NUMA locality during page-table walks. Bio: Reto Achermann is a Postdoctoral Research Fellow at the Systopia Lab of the Computer Science Department at the University of British Columbia working with Prof. Margo Seltzer. His research interests are at the intersection of memory and storage systems, hardware platforms, formal specification, and software synthesis. He holds a PhD in computer science from ETH Zurich where he was advised by Prof. Timothy Roscoe. He was part of the Barrelfish research operating system team, working on various subsystems such as memory management, capabilities, hardware abstractions, device drivers. During his PhD, Reto Achermann worked on Sockeye, a framework to formally specify hardware as seen by software; CleanQ, a queue interface with formally specified semantics, and the Enzian platform, a research computer combining a Cavium ThunderX processor and a large FGPA connected by the native coherence interconnect. During his PhD he also did research internships at VMware Research and Hewlett-Packard Labs. Jun 08 2021 16.00 - 17.00 ICSA Colloquium Talk - 08/06/2021 Reto Achermann (University of British Columbia) Zoom
ICSA Colloquium Talk - 08/06/2021 Title: Fast Local Page Tables for NUMA Servers with Mitosis Abstract: Increasing memory heterogeneity mandates careful data placement to hide the non-uniform memory access (NUMA) effects on applications. While NUMA optimizations have focused on application data for decades, they have ignored the placement of kernel data structures due to their small memory footprint. Yet, misplacement of kernel data structures in large NUMA systems can lead to significant slowdown of application performance. In this talk, I am going to demonstrate the adverse effect of misplaced kernel data structures at the example of page tables. To quantify the possible slow down, we conduct a systematic study of various allocation policies for both, the data pages and page-table pages, and its effects on the application's performance. Moreover, we show that in a multi-threaded application, memory accesses to remote NUMA nodes in during to page-table walks are inevitable. Based on our findings, we present Mitosis a set of kernel and hypervisor extensions that eliminate remote memory accesses during page-table walks. Mitosis replicates and/or migrates the application's page table across the NUMA nodes of the system. With Mitosis, applications observe a significant speedup due to improved NUMA locality during page-table walks. Bio: Reto Achermann is a Postdoctoral Research Fellow at the Systopia Lab of the Computer Science Department at the University of British Columbia working with Prof. Margo Seltzer. His research interests are at the intersection of memory and storage systems, hardware platforms, formal specification, and software synthesis. He holds a PhD in computer science from ETH Zurich where he was advised by Prof. Timothy Roscoe. He was part of the Barrelfish research operating system team, working on various subsystems such as memory management, capabilities, hardware abstractions, device drivers. During his PhD, Reto Achermann worked on Sockeye, a framework to formally specify hardware as seen by software; CleanQ, a queue interface with formally specified semantics, and the Enzian platform, a research computer combining a Cavium ThunderX processor and a large FGPA connected by the native coherence interconnect. During his PhD he also did research internships at VMware Research and Hewlett-Packard Labs. Jun 08 2021 16.00 - 17.00 ICSA Colloquium Talk - 08/06/2021 Reto Achermann (University of British Columbia) Zoom
Jun 08 2021 16.00 - 17.00 ICSA Colloquium Talk - 08/06/2021 Reto Achermann (University of British Columbia)