FDP WAF Stress Result Summary: R0 Window-Local 32GiB 3-Mode Batch

작성일: 2026-06-15 기준 시나리오: fdp/reports/2026-06-15_poc_harness_evaluation_scenario.md Run output: /home/ny/fdp_measure_waf_r0_window32g_readme8

Overall Result

PASS. no_fdp, mixed, separated 3모드 모두 R0 window-local 32GiB precondition 후 README 기준 stress, warmup=2, measurement=8 조건으로 완료됐고, 세 모드 모두 WAF가 available로 산출됐다.

Metric	Value
Result status	PASS
Mode(s)	no_fdp, mixed, separated
Run ID	R_no_fdp_20260615T222219, R_mixed_20260615T223842, R_separated_20260615T225510
Device path	/dev/ng0n1
Block device path	/dev/nvme0n1
Workers	9 per mode
Warmup iterations	2
Measurement iterations	8
Total replay processes	90 per mode
Process failures	0
Replay record failures	0
Operation errors	0
Total wall time	no_fdp 599.24s, mixed 605.94s, separated 597.14s
Warmup wall time	no_fdp 118.13s, mixed 118.68s, separated 117.74s
Measurement wall time	no_fdp 481.10s, mixed 487.25s, separated 479.39s

Precondition

원문 PoC 기준의 clean-state WAF artifact를 피하기 위해 각 mode 직전에 window-local R0 precondition을 수행했다.

Metric	Value
Precondition status	used
Precondition type	window-local
R0 output	/home/ny/fdp_measure_waf_r0_window32g_readme8/R0_precondition_20260615T222201, /home/ny/fdp_measure_waf_r0_window32g_readme8/R0_precondition_20260615T223825, /home/ny/fdp_measure_waf_r0_window32g_readme8/R0_precondition_20260615T225452
Namespace discard	yes
Offset bytes	2199023255552
Size	32GiB
Sequential fill	yes
Random overwrite	yes
Random passes	1
Vendor media counter during R0	collected

WAF Result

Mode	Host write bytes delta	Host write delta	Media write bytes delta	Media write delta	WAF	Write amplification over host writes	WAF status
no_fdp	511,058,432,000	475.96GiB	519,170,949,120	483.52GiB	1.015874	1.587%	available
mixed	500,674,560,000	466.29GiB	508,619,390,976	473.69GiB	1.015868	1.587%	available
separated	373,750,784,000	348.08GiB	379,687,010,304	353.61GiB	1.015883	1.588%	available

separated는 host/media write volume이 가장 낮았다. 다만 WAF ratio 자체는 세 모드가 모두 약 1.0159로 거의 같아서, 이번 batch만으로 "separated가 WAF ratio를 낮췄다"는 결론은 약하다.

Expected vs Observed

원문 PoC와 README의 기대 방향은 lifetime이 비슷한 KV cache stream을 분리하면 SSD 내부 GC의 불필요한 copy가 줄고, 그 결과 WAF와 NAND wear가 감소하며 p99 tail latency가 안정화된다는 것이다. README 기준으로는 mixed가 의도적으로 나쁜 FDP placement이고, separated가 lifetime/RUH 분리 효과를 확인하는 모드다.

기대 항목	기대 방향	이번 관측값	판단
WAF ratio	separated가 no_fdp/mixed보다 낮아지는 방향	no_fdp 1.015874, mixed 1.015868, separated 1.015883	미입증. 세 모드 WAF ratio가 사실상 동일
Host write pressure	separated가 baseline 대비 감소	separated는 no_fdp 대비 -26.9%, mixed 대비 -25.4%	긍정 신호
Media write pressure	separated가 media write를 줄이는 방향	separated 353.61GiB, no_fdp 483.52GiB, mixed 473.69GiB	긍정 신호
Tail latency	separated에서 p99가 안정화되는 방향	write p99: separated 2.398ms, no_fdp 2.732ms, mixed 3.337ms	긍정 신호
Clean-state caveat	WAF 평가 전에 preconditioning 적용	각 mode 전 32GiB window-local R0 수행	충족

따라서 이번 batch는 원문/README의 기대 중 write pressure 감소와 tail latency 개선에는 맞는 신호를 보였다. 반면 핵심 지표인 WAF ratio 자체의 감소는 이번 결과에서 확인되지 않았다. 이 결과는 "separated가 WAF를 낮췄다"가 아니라, "separated가 이번 workload에서 host/media write volume과 p99를 낮췄지만 WAF ratio는 동일 수준이었다"로 해석해야 한다.

Counter Collection

Counter	Source	Status
Host writes	nvme smart-log /dev/nvme0n1 -o json	collected
Media writes	nvme ocp smart-add-log /dev/nvme0n1 -o json	collected
FDP stats	outer nvme fdp stats /dev/nvme0 --endgrp-id=1 snapshots	collected
FDP RUH logs	xNVMe FDP log commands	not collected

Notes:

• measurement_after.json warnings contain xnvme unavailable; skipped FDP logs.
• Formal WAF used vendor OCP Physical media units written, not FDP MBMW proxy.

Measurement Workload Pressure

The table below uses the separated mode workload mapping. Measurement-only logical store writes are shown for measurement_iterations=8.

Workload	Workers	Class	FDP Data RUHs	FDP Metadata RUHs	Logical Store Writes	Store records	Prefetch records
llama8b_chat_chunk256	2	hot_churn	0,1	2	160GiB	40,960	5,120
llama70b_longctx_chunk1024	1	large_model	3,4	5	80GiB	2,560	2,512
rag_shared_prefix_chunk512	2	cold_rag	3,4	5	192GiB	12,288	48,864
random_prompts_chunk128	3	hot_churn	0,1	2	192GiB	98,304	1,536
metadata_heavy_small_objects	1	metadata_heavy	6	7	64GiB	65,536	0
Total	9				688GiB	219,648	58,032

Latency

Mode	write p50 ms	write p99 ms	read p50 ms	read p99 ms	all p99 ms
no_fdp	0.083	2.732	N/A	0.319	2.672
mixed	0.084	3.337	N/A	0.308	3.234
separated	0.079	2.398	N/A	0.303	2.338

separated가 이번 batch에서 write p99와 all p99 모두 가장 낮았다.

Log Observations

Replay completed cleanly. The warning categories that appeared in the earlier probe runs did not recur in this batch.

Warning Category	Count	Files
Cleaning up in-flight store task	0	0
Failed batched memory allocation	0	0
Device metadata slot_bytes mismatch	0	0
FDP metadata payload rejected	0	0
io_uring fixed-buffer registration disabled	0	0
Tracebacks	0	0
ERROR logs	0	0
Exceptions	0	0

Validity Against Scenario

Criterion	Required by scenario	Observed	Status
Original goal: lifetime separation	separated RUH mapping separates hot/cold/metadata	hot/small/random data 0,1, cold/large/rag data 3,4, metadata 2/5/7	PASS
Original goal: WAF available	host and media counters collected	WAF available in all 3 modes	PASS
Original goal: clean-state caveat addressed	R0 precondition status documented	32GiB window-local R0 before each mode	PASS
README scale	stress	stress	PASS
README warmup	2	2	PASS
README measurement	8	8	PASS
Process failures	0	0	PASS
Replay record failures	0	0	PASS
Host write delta	> 0	all modes > 0	PASS
Media write delta	> 0	all modes > 0	PASS
p99 latency available	write p99 at minimum	write/all p99 available; read p99 parsed from read-like records	PASS

Conclusion

This batch confirms that the PoC harness can run the README-style stress, warmup=2, measurement=8 workload under the original-document requirement of preconditioning before WAF evaluation. All three modes completed without replay failures, and formal WAF was available in all modes.

The main positive signal is that separated reduced host write volume by 26.9% versus no_fdp and 25.4% versus mixed, while also showing the lowest write p99. The WAF ratio itself was nearly identical across all modes at about 1.0159, so this run supports reduced write pressure and tail latency for separated, but not a strong WAF-ratio improvement claim. The next step is either repeated trials under the same 32GiB R0 condition or an expanded R0 span of 256GiB/512GiB to test whether the result is stable.