FDP PoC Stage 4 Summary

Historical note: this file summarizes the earlier warmup=0, measurement=1 Stage 4 smoke/diagnostic runs. The README-aligned formal scenario is now documented in runbook.md and ../reports/2026-06-15_poc_harness_evaluation_scenario.md: TRACE_SCALE=stress, WARMUP_ITERATIONS=2, MEASUREMENT_ITERATIONS=8, mode order no_fdp -> mixed -> separated, and report.md linking each output summary.md.

Current status

• Target: /dev/nvme0n1 block namespace, /dev/ng0n1 raw generic namespace, controller /dev/nvme0, endgrp 1.
• Stage 3 direct FDP sanity passed: write_uring ... dspec=0, readback matched.
• Stage 4 3-mode stress comparison is now valid under one common setup.
• Output root: /home/ny/fdp_measure_3mode_final
• Parsed summary was regenerated from parse_results.py because the first redirected results_summary.csv was empty.

Final 3-mode result

Parsed CSV:

run_dir,mode,host_write_bytes_delta,fdp_hbmw_delta,fdp_mbmw_delta,fdp_mbe_delta,write_p99_ms,all_p99_ms
R_mixed_20260615T084519,mixed,43859968000,43860094976,44577980416,39862665216,1.2903,1.2211
R_no_fdp_20260615T083105,no_fdp,385622528000,385622347776,399353851904,418557984768,1.3531,1.3495
R_separated_20260615T083813,separated,198493184000,198492884992,202153451520,199313326080,3.3568,3.2733

Per-mode summary:

Mode	Run dir	Workers	Records	Failures	host write delta	FDP HBMW delta	write p99 ms	all p99 ms	WAF status
no_fdp	R_no_fdp_20260615T083105	9	27,240	0	385,622,528,000	385,622,347,776	1.353	1.349	unavailable: media/NAND write counter missing
separated	R_separated_20260615T083813	9	27,240	0	198,493,184,000	198,492,884,992	3.357	3.273	unavailable: media/NAND write counter missing
mixed	R_mixed_20260615T084519	9	27,240	0	43,859,968,000	43,860,094,976	1.290	1.221	unavailable: media/NAND write counter missing

Interpretation:

• This run is the first valid same-condition 3-mode comparison in this PoC stage.
• separated is no longer stuck at host_write_bytes_delta=0; the earlier 0 result was caused by mode-specific raw-device window shift.
• In this workload and measurement setup, host writes rank as no_fdp > separated > mixed.
• The original hypothesis "separated should be lowest" is not supported by this run.
• WAF itself is still unavailable because there is no vendor media/NAND write counter.

Counter consistency

Internal measurement and outer SMART/FDP snapshots agree.

Mode	Internal host delta	Outer SMART delta	Outer FDP HBMW delta
no_fdp	385,622,528,000	385,622,528,000	385,622,347,776
separated	198,493,184,000	198,493,184,000	198,492,884,992
mixed	43,859,968,000	43,859,968,000	43,860,094,976

This means:

• harness SMART parsing is working,
• after-measurement settle timing is sufficient,
• and the current result set is usable for comparison.

Next step

This is still a single valid comparison set. The next required step is repeated trials under the same condition to measure run-to-run variance.

• wrapper: run_repeated_trials.sh
• aggregate parser: summarize_repeated_trials.py
• recommended baseline: REPEATS=3, ORDER_STRATEGY=rotate, TRACE_SCALE=stress

Validated mode mapping

Mode	FDP	RUH policy
no_fdp	off	no FDP placement hints
separated	on	hot/small/random data [0,1] metadata [2]; cold/rag/large data [3,4] metadata [5]; metadata-heavy data [6] metadata [7]
mixed	on	all workers share data [0,1,3,4,6,7], metadata [2,5]

Next command: actual smoke without R0

This does not run blkdiscard, but it still writes to the raw NVMe namespace. Run it only on the confirmed FDP test SSD.

cd /home/ny/work/fdp/04_measure

for mode in no_fdp mixed separated; do
  sudo env \
    HOME=/home/ny \
    REPO=/home/ny/work/LMCache-fdp-poc-src \
    NVME_DEV=/dev/nvme0 \
    NVME_NS=/dev/nvme0n1 \
    NVME_IO_DEV=/dev/ng0n1 \
    FDP_ENDGRP_ID=1 \
    FDP_RUH_COUNT=8 \
    TRACE_SCALE=smoke \
    WARMUP_ITERATIONS=0 \
    MEASUREMENT_ITERATIONS=1 \
    OUT=/home/ny/fdp_measure_retry \
    bash run_poc_harness.sh "$mode"
done

/usr/bin/python3 parse_results.py /home/ny/fdp_measure_retry > /home/ny/fdp_measure_retry/results_summary.csv
/usr/bin/python3 write_report.py /home/ny/fdp_measure_retry > /home/ny/fdp_measure_retry/report.md

Optional command: R0-normalized comparison

This runs blkdiscard /dev/nvme0n1 before each mode and is destructive.

cd /home/ny/work/fdp/04_measure

for mode in no_fdp mixed separated; do
  sudo env \
    HOME=/home/ny \
    NVME_DEV=/dev/nvme0 \
    NVME_NS=/dev/nvme0n1 \
    FDP_ENDGRP_ID=1 \
    OUT=/home/ny/fdp_measure_retry \
    PRECONDITION_SIZE_GB=16 \
    bash run_R0_precondition.sh

  sudo env \
    HOME=/home/ny \
    REPO=/home/ny/work/LMCache-fdp-poc-src \
    NVME_DEV=/dev/nvme0 \
    NVME_NS=/dev/nvme0n1 \
    NVME_IO_DEV=/dev/ng0n1 \
    FDP_ENDGRP_ID=1 \
    FDP_RUH_COUNT=8 \
    TRACE_SCALE=smoke \
    WARMUP_ITERATIONS=0 \
    MEASUREMENT_ITERATIONS=1 \
    OUT=/home/ny/fdp_measure_retry \
    bash run_poc_harness.sh "$mode"
done

/usr/bin/python3 parse_results.py /home/ny/fdp_measure_retry > /home/ny/fdp_measure_retry/results_summary.csv
/usr/bin/python3 write_report.py /home/ny/fdp_measure_retry > /home/ny/fdp_measure_retry/report.md

Interpretation

The current runs prove more than harness execution: they provide a valid same-condition 3-mode comparison. The next work item is not rerunning the same baseline, but explaining why mixed produced the lowest host-write delta and whether that reflects trace shape, RUH policy, or an artifact of this PoC harness.