Gahow Wang
547611e022
D2: run_benchmark.sh and run_experiments.sh still pass --time-scale and --max-inflight-sessions to the replayer, but those flags were removed when the project moved to trace-driven dispatch. The scripts cannot run as-is. D3: ~25 ad-hoc analyze_* / compare_* / profile_* / final_* scripts and a handful of single-experiment run_*.sh point at /home/admin/cpfs paths, deleted output directories, or a sampled trace file that no longer exists. Keep them in scripts/legacy/ for historical reference; the scripts that remain in scripts/ (analyze_trace, analyze_breakdown, analyze_cache_hit, analyze_eviction, compare_results, compute_roofline, sample_trace, analyze_agentic_patterns, simulate_cache_policies, plus launch_*.sh, gpu_monitor.sh, bench.sh) cover the current workflow. Adds scripts/legacy/README.md to document the archival policy.
87 lines
3.8 KiB
Python
87 lines
3.8 KiB
Python
"""Final comparison of PD-Combined vs PD-Separated (Mooncake/RDMA)."""
|
|
import json, statistics, os
|
|
|
|
def pct(vals, q):
|
|
return vals[min(int(q * len(vals)), len(vals) - 1)] if vals else 0
|
|
|
|
# Combined (16 sessions) - completed run
|
|
rows_c = [json.loads(l) for l in open("outputs/v18_combined_1000req/metrics.jsonl")]
|
|
ok_c = [r for r in rows_c if not r.get("error")]
|
|
ttfts_c = sorted([r["ttft_s"] for r in ok_c if r.get("ttft_s")])
|
|
tpots_c = sorted([r["tpot_s"] for r in ok_c if r.get("tpot_s") and r["tpot_s"] > 0])
|
|
lats_c = sorted([r["latency_s"] for r in ok_c if r.get("latency_s")])
|
|
sc = json.load(open("outputs/v18_combined_1000req/metrics.summary.json"))
|
|
|
|
# PD-Separated Mooncake (first 200 stable requests)
|
|
rows_d = [json.loads(l) for l in open("outputs/v18_pd_mooncake_lowconc/metrics.jsonl")][:200]
|
|
ok_d = [r for r in rows_d if not r.get("error")]
|
|
ttfts_d = sorted([r["ttft_s"] for r in ok_d if r.get("ttft_s")])
|
|
tpots_d = sorted([r["tpot_s"] for r in ok_d if r.get("tpot_s") and r["tpot_s"] > 0])
|
|
lats_d = sorted([r["latency_s"] for r in ok_d if r.get("latency_s")])
|
|
|
|
sep = "=" * 70
|
|
print(sep)
|
|
print(" PD-Combined vs PD-Separated (Mooncake/RDMA)")
|
|
print(" vLLM 0.18.1 | Qwen3-Coder-30B-A3B | 8xH20")
|
|
print(sep)
|
|
|
|
header = " {:<12} {:>16} {:>16} {:>10}".format(
|
|
"Metric", "Combined(TP=8)", "PD-Sep(TP=4+4)", "Delta")
|
|
print(header)
|
|
dash = " {:<12} {:>16} {:>16} {:>10}".format("-" * 12, "-" * 16, "-" * 16, "-" * 10)
|
|
print(dash)
|
|
|
|
req_c = "{}/{}".format(len(ok_c), len(rows_c))
|
|
req_d = "{}/{}".format(len(ok_d), len(rows_d))
|
|
print(" {:<12} {:>16} {:>16}".format("Requests", req_c, req_d))
|
|
|
|
data = [
|
|
("TTFT p50", pct(ttfts_c, 0.5), pct(ttfts_d, 0.5)),
|
|
("TTFT p90", pct(ttfts_c, 0.9), pct(ttfts_d, 0.9)),
|
|
("TPOT p50", pct(tpots_c, 0.5), pct(tpots_d, 0.5)),
|
|
("TPOT p90", pct(tpots_c, 0.9), pct(tpots_d, 0.9)),
|
|
("E2E p50", pct(lats_c, 0.5), pct(lats_d, 0.5)),
|
|
("E2E p90", pct(lats_c, 0.9), pct(lats_d, 0.9)),
|
|
]
|
|
|
|
for label, cv, dv in data:
|
|
delta = "{:+.0f}%".format((dv / cv - 1) * 100) if cv > 0 else "N/A"
|
|
print(" {:<12} {:>15.3f}s {:>15.3f}s {:>10}".format(label, cv, dv, delta))
|
|
|
|
cache_c = sc.get("prefix_cache_hit_ratio", 0)
|
|
print(" {:<12} {:>15.1f}% {:>16}".format("Cache hit", cache_c * 100, "N/A"))
|
|
tput_c = len(ok_c) / sc.get("wall_clock_s", 1)
|
|
print(" {:<12} {:>14.2f}/s {:>16}".format("Throughput", tput_c, "~0.06/s"))
|
|
|
|
print()
|
|
print(sep)
|
|
print(" CONCLUSIONS FOR AGENTIC WORKLOAD")
|
|
print(sep)
|
|
print()
|
|
print(" Trace characteristics:")
|
|
print(" - I/O ratio: 61.5x (strongly prefill-dominated)")
|
|
print(" - 39% requests > 32k input tokens")
|
|
print(" - 16% prefix block sharing across sessions")
|
|
print(" - 53% prefix cache hit ratio (APC)")
|
|
print()
|
|
print(" PD separation findings:")
|
|
|
|
delta_tpot = (pct(tpots_d, 0.5) / pct(tpots_c, 0.5) - 1) * 100 if tpots_c else 0
|
|
delta_ttft = (pct(ttfts_d, 0.5) / pct(ttfts_c, 0.5) - 1) * 100 if ttfts_c else 0
|
|
delta_e2e = (pct(lats_d, 0.5) / pct(lats_c, 0.5) - 1) * 100 if lats_c else 0
|
|
|
|
print(" 1. TPOT {:+.0f}% - decode isolation benefit is {}".format(
|
|
delta_tpot, "marginal" if abs(delta_tpot) < 20 else "significant"))
|
|
print(" 2. TTFT {:+.0f}% - KV transfer + TP=4 overhead dominates".format(delta_ttft))
|
|
print(" 3. E2E {:+.0f}% - net negative on single-machine".format(delta_e2e))
|
|
print(" 4. Stability: Mooncake connector crashes after ~200 reqs under load")
|
|
print()
|
|
print(" Recommendation:")
|
|
print(" - Single-machine 8 GPU: Combined mode is better (lower TTFT, stable)")
|
|
print(" - Multi-machine: PD-Sep is promising IF cross-machine latency")
|
|
print(" is hidden by RDMA and prefill doesn't share GPU with decode")
|
|
print(" - Key bottleneck: this workload's heavy prefill (avg 32k tokens)")
|
|
print(" makes KV transfer cost non-trivial relative to prefill time")
|
|
print(" - Prefill-as-a-Service (Goal 5) should focus on cross-machine")
|
|
print(" KV cache sharing, not same-machine PD split")
|