This work focuses on the bottleneck of file-mapping in NVM filesystems.

The “file mapping” here aims the translation and allocation from logical address (inode-num, logical block) to physical address in I/O path. They claim that file mapping comprise up to 70% of all overhead.

related works: strata, ZoFS, NOVA, SplitFS…

Their works:

1. analyze file mapping in PM
   • lots of experiments in this part, including page cache, size, space utilization, fragmentation…
   • Also legacy structures suffered perf loss on large files and update ops.

     This part contains lots of details!

2. optimize file mapping structures
   • cuckoo hashing
     • …
   • HashFS
     • use hash table with linear probing to index file data. and the file offset is the hashing collision offset.
     • intuition behind this design: since page cache is harmful, hash table can make sparse and random update operations efficient. As a result, it also saves cost of block allocator management.
     • the implementation is only depend on memory primitives
     • To avoid resize hash table, statically allocate a super big global table.

       or can we use multiple tables?

     • Use SIMD to accelerate I/O for lookup, since lookup can be done in parallel.
3. evaluate them on real workloads
   • YCSB: 10-45% gains
   • filebench: …
   • locality: > ???

A good work lead by findings of experiments.

refer

1. Neal, Ian, et al. “Rethinking File Mapping for Persistent Memory.” 19th {USENIX} Conference on File and Storage Technologies ({FAST} 21). 2021.