Memory-mapped file
Also known as file-backed mapping
In plain English
A memory-mapped file exposes file contents through memory addresses while the operating system manages which pages are resident.
Technical definition
A memory-mapped file maps file-backed data into a process address space, allowing the operating system to load and reclaim pages as they are accessed.
Engineering details
The article replaces the Engram allocation with local SSD-backed mappings. Pages already resident in the filesystem cache can satisfy reads without another SSD access. File backing lets the operating system reclaim those pages when other applications need RAM.
Why it matters
The mapped file size is not the same as physical RAM use or bytes read from the SSD during a benchmark. Cache residency and memory pressure must be reported to understand what an experiment actually measures.
How to read it in InferenceX
The B200 experiment uses CPU row-ID handling, deduplication, pinned-buffer gathering, and GPU transfers around graph segments. It did not enable GDS. Consequently its measured gap from DRAM includes the implementation overhead as well as any physical storage activity.