Gahow Wang 2fee355626 Adaptive v2 (selective Mooncake offload): worse than baseline ...

Implemented --offload mode: HEAVY requests (>20k new tokens) get P on
least-loaded instance, KV via Mooncake RDMA, D on session-sticky instance.
WARM/MEDIUM stay co-located (no KV transfer). All 8 instances run kv_both.

Result (200 req, same instances, fresh restart):
  Baseline (no offload):   TTFT=1.073  TPOT90=0.074  E2E=5.086
  Offload HEAVY:            TTFT=1.462  TPOT90=0.077  E2E=6.847
  Delta:                    +36%        +4%            +35%

Conclusion: even selective KV transfer (only 44% of requests) adds more
overhead than the isolation benefit provides. On single-machine 8 GPU,
PD-combined with hybrid routing is strictly optimal. No form of KV
transfer — full PD-sep, selective offload, or otherwise — improves
over co-located serving for this workload.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-05-22 10:14:10 +08:00

..

ab_gpu_test.sh

Add GPU utilization A/B test and fix cache-aware proxy bugs

2026-05-21 22:13:38 +08:00

analyze_3way.py

Add P/D ratio ablation: 6P+2D vs 4P+4D vs Combined

2026-05-21 22:42:20 +08:00

analyze_ablations.py

Ablation 2: fire-and-forget vs await-prefill scheduling

2026-05-21 23:02:42 +08:00

analyze_agentic_patterns.py

Balanced session-sticky routing + agentic workload pattern analysis

2026-05-22 01:50:27 +08:00

analyze_breakdown.py

Add per-request breakdown profiling, identify KV cache memory bottleneck

2026-05-22 00:13:50 +08:00

analyze_cache_hit.py

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

analyze_eviction.py

KV cache lifecycle design + eviction loss analysis

2026-05-22 01:27:22 +08:00

analyze_gpu_ab.py

Add GPU utilization A/B test and fix cache-aware proxy bugs

2026-05-21 22:13:38 +08:00

analyze_trace.py

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

cache_aware_proxy.py

Adaptive v2 (selective Mooncake offload): worse than baseline

2026-05-22 10:14:10 +08:00

compare_adaptive.py

Adaptive prefill offload v1: implementation + experiment

2026-05-22 01:00:10 +08:00

compare_balanced.py

Balanced routing result: APC +4pp but latency +23% (cache-load tradeoff)

2026-05-22 02:13:15 +08:00

compare_results.py

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

compute_roofline.py

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

final_comparison.py

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

final_gpu_comparison.py

Final GPU comparison: hybrid routing matches baseline latency with better APC

2026-05-22 03:14:05 +08:00

gpu_monitor.sh

Add GPU utilization A/B test and fix cache-aware proxy bugs

2026-05-21 22:13:38 +08:00

launch_pd_mooncake.sh

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

launch_pd_separated.sh

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

launch_vllm.sh

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

profile_fnf.py

Add per-request breakdown profiling, identify KV cache memory bottleneck

2026-05-22 00:13:50 +08:00

run_benchmark.sh

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

run_experiments.sh

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

sample_trace.py

Agentic workload PD separation analysis with trace-driven benchmarks

2026-05-21 21:21:57 +08:00

simulate_cache_policies.py

Cache policy simulation: routing quality dominates, not eviction policy

2026-05-22 01:28:53 +08:00