Tiered vs. flat: where does a block live?
Toggle the architecture. Flat = decide medium at write time, never migrate.
Why delivered hit rate matters
A block on disk that arrives late is recomputed anyway. Step through a request.
Premise 1 fails: DRAM does not filter cold data
Reported in the paper's Table 1 (as quoted in the episode).
Hover a bar.
Premise 2 fails: staging and scatter
Session blocks on the drive. Illustrative block placement, not measured data.
Causal chain behind the collapse
As described by the authors; click a link.
Click a box.
GPUDirect Storage: transport vs. layout
GDS removes the bounce buffer. An NVMe read still needs a contiguous range, so scattered blocks still mean many small commands.
FC-GVA: placement at write time
Toggle signals. Rule order follows the episode's description. Thresholds tau and L_agent: no values published.
64-bit block address
Hover segments. One hash lookup returns the address; medium bits pick DRAM copy vs SSD read.
Hover a field.
FC-CIO: session-contiguous, layer-major extent
Toggle read pattern. Gaps under 64 KiB are bridged. Illustrative block counts.
FC-MAS: two queues
Results on the Claude Code trace
8×B200, 1 TiB DRAM, 4 TiB SSD per system, SGLang, DeepSeek-V4-Flash, ~99,000 requests / 393 sessions, 0.1–1.0 req/s. Reported values only.
Ablation: remove FC-MAS
Only component ablated on the agentic trace: about 10–11% worse mean TTFT, 27% worse P99.
What the experiments test
Coverage matrix from the episode's critique. Hover a cell.
Hover a cell.
Placement risk: a wrong choice sticks
Step a session through time. Illustrative scenario.