Efficient and Correct Persistent Memory File Systems

Persistent memory (PM) is a novel storage technology that enables byte-granular direct access from the CPU with low latency.
Compared to traditional asynchronous block storage, this access paradigm allows file systems to offer stronger persistence guarantees at lower latency.
However, it also introduces new challenges for performance, efficiency, and correctness.
PM's small atomic write size requires careful use of PM primitives to prevent data loss in the event of a crash.
Overloading PM with parallel accesses results in expensive CPU stalls.

This thesis investigates the efficiency and correctness of PM file systems.
First, we introduce efficiency metrics that quantify CPU time and energy cost per unit of storage access.
We show that many existing PM file systems perform poorly under parallel load.
To address PM overload, we design mitigation mechanisms that integrate with existing file systems and a monitoring technique to attribute direct-access PM traffic to processes.

Second, to improve correctness, we present Suvi, an approach to black-box crash consistency testing for PM file systems.
Suvi traces a file system's PM and NVMe SSD accesses in a virtual machine and replays the trace with an accurate simulation of x86 store-order semantics.
... mehr
Suvi generates crash images using two heuristics to avoid combinatorial explosion, and then automatically analyzes the crash images to detect atomicity bugs.

Together, these contributions provide measurement tools, mitigation strategies, and testing infrastructure to make PM file systems more efficient and more reliable.