File size: 7,298 Bytes
23a7a20
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
---
license: mit
task_categories:
- other
pretty_name: Stream2LLM Data
tags:
- llm-serving
- streaming-inference
- vllm
- systems
viewer: false
---

# Stream2LLM Dataset

Dataset for the paper: [Stream2LLM: Overlap Context Streaming and Prefill for Reduced Time-to-First-Token (TTFT)](https://huggingface.co/papers/2604.16395) (MLSys 2026 artifact evaluation).

GitHub Repository: [https://github.com/rajveerb/stream2llm/](https://github.com/rajveerb/stream2llm/)

Contains workload traces, experiment run logs, and performance model measurements used to produce all figures, tables, and inline numbers in the paper.

This repository is a git submodule of the main [Stream2LLM artifact](https://github.com/rajveerb/stream2llm/tree/mlsys_artifact) (branch: `mlsys_artifact`). Data files are stored with Git LFS on HuggingFace. To fetch everything:

```bash
# Option 1: Clone parent repo with submodules, then pull LFS files
git clone --recurse-submodules -b mlsys_artifact https://github.com/rajveerb/stream2llm.git
cd stream2llm/data && git lfs install && git lfs pull && cd ../..

# Option 2: If you already cloned without submodules
git submodule update --init
cd data && git lfs install && git lfs pull && cd ..
```

## Directory Structure

```
data/
├── anns/                          # ANNS workload data
│   ├── res/                       # 4,997 pipeline trace CSVs
│   ├── retrieved_corpus_content.*.json  # Corpus content shards
│   ├── query_trace_map_5k.json    # Query-to-trace mapping
│   └── compute_workload_stats.py  # Workload statistics script
├── crawl/                         # Crawler workload data
│   ├── traces/simpleQA_ALL/       # 4,322 query trace CSVs
│   └── compute_workload_stats.py  # Workload statistics script
├── perf_model/                    # Performance model measurements
│   ├── recomputation/             # 7 recomputation latency JSONs
│   └── swap/                      # 11 swap latency JSONs
└── run_log/                       # Experiment run logs
    ├── crawler/                   # 5 crawler experiment configurations
    └── anns/                      # 5 ANNS experiment configurations
```

## Workload Traces

### ANNS (`anns/res/`)

4,997 pipeline trace files from approximate nearest neighbor search workloads. Each file is named `_L10000_W8_query<ID>_pipeline_trace.csv` and contains:

| Column | Description |
|--------|-------------|
| `StartTime_us` | Chunk start time in microseconds |
| `EndTime_us` | Chunk end time in microseconds |
| `StartIteration` | Starting iteration index |
| `EndIteration` | Ending iteration index |
| `PipelinePool` | Tuple of candidate IDs in the pipeline pool |

Each row represents a chunk arrival — a batch of ANNS iterations that produces new candidate results. Queries have 1–26 chunks (median 4), with inter-chunk arrival times ranging from sub-millisecond to ~9 seconds (median 37 ms).

### Crawler (`crawl/traces/simpleQA_ALL/`)

4,322 query trace files from a web crawling workload (SimpleQA question-answering). Each file is named `query_<ID>.csv` and contains:

| Column | Description |
|--------|-------------|
| `type` | Event type: `tavily_search` or `page_scrape` |
| `startTime` | Event start time in seconds |
| `endTime` | Event end time in seconds |
| `query` | The original query string (on search rows) |
| `links_found` | Number of links returned by search |
| `url` | URL scraped (on page_scrape rows) |
| `content_length` | Length of scraped content in characters |
| `content` | Scraped page text |
| `link_idx` | Index of the link being scraped |

Each row is a chunk event. The first row is typically a `tavily_search` followed by `page_scrape` events. Queries have 1–17 chunks (median 8), with inter-chunk arrival times of 3 ms to ~35 seconds (median 701 ms).

## Performance Model (`perf_model/`)

Microbenchmark measurements for KV cache eviction cost modeling, collected across multiple GPU types (A40, A100, H100, H200) and model configurations (8B, 70B with varying tensor parallelism).

- **`recomputation/`**: Recomputation latency in ms, keyed by number of tokens recomputed (e.g., 16, 32, ..., 8192). Each key maps to an array of repeated measurements.
- **`swap/`**: Swap (CPU↔GPU transfer) latency in ms, same key structure. Files ending in `_kernel_latency.json` contain kernel-only timings; others include end-to-end transfer overhead.

Hardware configurations: `A40`, `A100`, `H100_tp2`, `H100_tp4_70B`, `H200_tp2`, `H200_tp4_70B`.

## Run Logs (`run_log/`)

Experiment outputs from the Stream2LLM serving system, organized by workload and configuration.

### Structure

```
run_log/<workload>/<config>/<scheduler>/<timestamp>/
    ├── config_<timestamp>.yaml   # Experiment configuration
    └── run_metrics.csv           # Per-event metrics log
```

### Configurations

| Workload | Config Directory | Description |
|----------|-----------------|-------------|
| Crawler | `H200_enhanced_schedulers_v1_full` | Standard H200 runs |
| Crawler | `H200_enhanced_schedulers_v1_full_delay_10` | With 10ms artificial delay (memory pressure) |
| Crawler | `H200_..._delay_10_recomp_only` | Delay + recomputation-only eviction |
| Crawler | `H200_..._delay_10_swap_only` | Delay + swap-only eviction |
| Crawler | `H100_enhanced_schedulers_v1_full` | Standard H100 runs |
| ANNS | `H200_enhanced_schedulers_v1_full` | Standard H200 runs |
| ANNS | `H200_enhanced_schedulers_v1_500q_delay_30` | With 30ms artificial delay (memory pressure) |
| ANNS | `H200_..._delay_30_recomp_only` | Delay + recomputation-only eviction |
| ANNS | `H200_..._delay_30_swap_only` | Delay + swap-only eviction |
| ANNS | `H100_enhanced_schedulers_v1_full` | Standard H100 runs |

### Schedulers

Each configuration contains results for four scheduling policies:

- **`default_vllm`**: Default vLLM scheduler (baseline)
- **`fcfs`**: First-come-first-served
- **`lcas`**: Last-chunk-arrival-stamp scheduler
- **`mcps`**: Most-chunks-processed scheduler

### Run Metrics CSV

The `run_metrics.csv` file logs timestamped events for each experiment run:

| Column | Description |
|--------|-------------|
| `event_timestamp` | Unix timestamp of the event |
| `event_type` | Event type (e.g., `replay_start`, `query_delay`, `chunk_sent`, `response_received`) |
| `query_id` | Query identifier |
| `request_id` | vLLM request identifier |
| `stream` | Whether streaming input was enabled |
| `concurrency` | Whether concurrent requests were enabled |
| `duration_secs` | Duration of the event in seconds |
| `details` | Additional event-specific information |
| `request_size` | Size of the request in tokens |
| `concurrent_requests` | Number of concurrent requests at event time |
| `replay_rate` | Poisson arrival rate used |
| `prev_event_type` | Previous event type for this query |

## Corpus Content (`anns/`)

The ANNS corpus is stored as sharded JSON files (`retrieved_corpus_content.*.json` and `retrieved_corpus_content.part.*.json`). Each file maps document IDs to their text content, used for constructing input sequences from ANNS retrieval results.

The `query_trace_map_5k.json` file maps query IDs to their corresponding pipeline trace filenames and query text.