Seminar • Systems and Networking • MemSnap: Persistent Memory without NVDIMMs using a Single Level Store

Please note: This half-hour seminar will take place in DC 1304.

Emil Tsalapatis, PhD candidate
David R. Cheriton School of Computer Science

Supervisor: Professor Ali Mashtizadeh

Software like databases and KV stores include significant amounts of userspace code to ensure correctness. These systems use file IO on top of conventional file systems because mmap() based persistence is difficult to implement correctly. File-based persistence however is prone to misuse and requires copying data from application memory to the disk during writes, lowering performance.

We claim that the cause of complexity in software like databases is the lack of persistence guarantees from the OS itself. As an alternative to existing file-based mechanisms we propose single level store (SLS) semantics that persist data without any user-level tracking. We build MemSnap, a database-oriented OS mechanism that removes the need for userspace writeahead logs or checkpointing. MemSnap instead tracks the working set using the virtual memory subsystem. MemSnap improves the performance of existing codebases by up to 4\x while removing code, allowing us to rethink the design of user space buffer caches and data management in general.

Bio: Emil Tsalapatis is a final-year PhD student working with Ali Mashtizadeh. His research spans operating systems, storage and cloud computing.

Please note: This is the first of two back-to-back Systems and Networking seminars.