~/talk

GeminiFS: A Companion File System for GPUs

Talkat FAST '25 · 2025-02-25..02-27
name:Shi Qiu
links:none on file
links:none on file
name:Yifan Hu
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
slides:none on file

GeminiFS embeds file metadata directly inside files so GPUs read address mappings without host roundtrips, extends the NVMe driver to map I/O queues into GPU memory so CPU and GPU set up control planes in parallel, and implements a GPU-friendly software page cache with warp-level locking and concurrent containers plus a programming library. Key components are embedded in-file metadata, NVMe IQ mapping to GPU memory, GPU-friendly page cache, libGemini API. Evaluation reports checkpoint latency down 65% versus host copy and 50% versus GDS and training runtime falling roughly 90% in their experiments.

AI-written summary · about

  1. 0:00Introduction
  2. 0:31Background and motivation
  3. 1:38Limitations of existing systems
  4. 3:41Design goals for GeminiFS
  5. 7:05GeminiFS design architecture
  6. 10:25Key implementation details
  7. 14:52Evaluation and conclusion
~/search
type to search 6,249 conferences, 138,987 speakers and 185,172 talks
~/help
cmdctrl k
search every conference, speaker and series
/
focus the filter search
h
home
c
conferences
s
speakers
e
series
t
light / dark — the status bar button does the same
?
this help
esc
close this, then clear every filter — or just the search, while you are typing in it

Ignored while you are typing. Pages without a filter rail have no search to focus.