prajwalpujari16/llm-serving-traces

⏱️ LLM Serving Traces & Scheduler Performance Benchmark

A curated empirical benchmark dataset containing real and synthetic LLM request arrival traces (conversational LMSYS, long-context RAG, code generation) and empirical scheduling policy comparisons (FCFS, SJF, Chunked Prefill, SLO-Aware, Disaggregated Serving) generated by the sched-sim discrete-event engine.

• GitHub Repository: Prajwal7-2002/sched-sim
• PyPI Engine: pip install sched-sim
• Documentation: prajwal7-2002.github.io/sched-sim
• Author: Prajwal Pujari (@Prajwal7-2002)

---

📊 Quick Start

Load the traces and scheduling benchmarks in Python with one line:

from datasets import load_dataset

# Load all three benchmark splits
dataset = load_dataset("prajwalpujari16/llm-serving-traces")

# Inspect master arrival traces
df_traces = dataset["serving_traces_master"].to_pandas()
print(df_traces[["request_id", "workload_category", "prompt_tokens", "output_tokens", "slo_ttft_ms"]])

# Inspect scheduler comparison leaderboard
df_sched = dataset["scheduler_benchmark_comparison"].to_pandas()
print(df_sched[["scheduler_policy", "workload_domain", "ttft_p99_ms", "goodput_pct"]])

---

🔬 Dataset Splits & Tables

1. serving_traces_master.csv

Contains arrival timestamps, prompt lengths, output lengths, prefill-to-decode ratios, and SLO latency thresholds across:

• Chatbot / Conversational: Dialogue requests from LMSYS Chatbot Arena with short prompts and moderate output tokens.
• Long-Context RAG: Document ingestion workloads with extreme prefill-to-decode ratios ($50:1$) causing severe prefill-decode interference.
• Code Generation / Autonomous Agents: Large output generation workloads with high memory consumption per request.
• Synthetic Poisson & Bursty Traffic: Standardized arrival processes ($\lambda = 10 \to 15\text{ req/s}$) for queueing theory and Little's Law validation.

2. scheduler_benchmark_comparison.csv

Empirical performance metrics across 5 production scheduling policies on NVIDIA A100 (80GB):

Scheduler Policy	Workload Domain	TTFT p50 (ms)	TTFT p99 (ms)	TPOT p50 (ms)	TPOT p99 (ms)	Goodput (%)
FCFS (Continuous Batching)	RAG	663.2	1704.5	32.2	69.5	40.0%
SJF (Shortest Prompt First)	RAG	663.2	1704.5	32.2	69.5	40.0%
Chunked Prefill (Budget: 512t)	RAG	625.0	1330.1	28.7	46.7	75.0%
SLO-Aware (Goodput Maximizer)	RAG	726.4	1704.5	31.3	50.0	47.5%
Disaggregated (DistServe)	RAG	663.2	1704.5	32.4	69.6	40.0%

3. hardware_validation_reference.csv

Empirical silicon ground-truth validation comparing discrete-event simulation predictions against physical GPU hardware runs (NVIDIA RTX 3050 Laptop GPU & AMD Ryzen 5 6600H) running Ollama (llama.cpp) and mini-infer (PyTorch eager).

---

🎯 Systems & Queueing Theory Insights

1. The Prefill-Decode Interference Problem:

   • In continuous batching, a 4,000-token RAG prefill request blocks active decode iterations. Decode requests experience a massive Time-Between-Tokens (TPOT) latency spike (from 28ms to >70ms).
   • Chunked Prefill splits large prefills into chunks of budget $C=512$, keeping active decode steps unstarved and boosting Goodput from 40.0% to 75.0%.

2. Little's Law in Serving Clusters:

   • Under steady-state arrival rate $\lambda$ and average latency $W$, the number of concurrent requests in flight is: $$L = \lambda \cdot W$$
   • For memory-bound KV-cache allocation, when $L \cdot \text{VRAM}{\text{req}} > \text{VRAM}{\text{GPU}}$, queue backpressure spikes exponentially.

---

🛠️ Reproducing Locally

# Install the zero-dependency discrete-event simulator
pip install sched-sim

# Run the 5-scheduler comparison on RAG traffic
sched-sim compare --workload rag --hardware a100

# Export Chromium / Perfetto trace for visual inspection
sched-sim run --scheduler chunked --trace-out trace.json

---

📜 Citation

@misc{pujari2026schedsim,
  author = {Prajwal Pujari},
  title = {LLM Serving Traces & Scheduler Performance Benchmark},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Data repository},
  howpublished = {\\url{https://huggingface.co/datasets/prajwalpujari16/llm-serving-traces}}
}