Skip to content

trace_format.md

Latest commit

 

History

History
66 lines (51 loc) · 2.43 KB

trace_format.md

File metadata and controls

66 lines (51 loc) · 2.43 KB
Raw

YieldOS-Lite Trace Format

YieldOS-Lite can replay CSV or JSONL traces. The replay adapter is intentionally small so public traces or production exports can be normalized without changing the simulator.

Required Fields

field type meaning
arrival_time_ms integer request arrival timestamp in milliseconds
prompt_tokens integer prompt length
output_tokens integer realized output length
tenant_id string tenant or workload owner
priority_class string interactive, standard, or batch

Optional Fields

field type default
request_id string generated from row number
ttft_slo_ms integer derived from priority class
itl_slo_ms integer derived from priority class
prefix_id string no shared prefix
abandoned_at_ms integer request is never abandoned
request_type string ignored by the simulator, kept for analysis

CSV Example

request_id,arrival_time_ms,prompt_tokens,output_tokens,tenant_id,priority_class,ttft_slo_ms,itl_slo_ms,prefix_id,abandoned_at_ms,request_type
r1,0,128,64,tenant-a,interactive,950,145,,,
r2,32,4096,180,tenant-b,standard,2600,280,doc-17,,rag_qa

JSONL Example

{"request_id":"r1","arrival_time_ms":0,"prompt_tokens":128,"output_tokens":64,"tenant_id":"tenant-a","priority_class":"interactive"}
{"request_id":"r2","arrival_time_ms":32,"prompt_tokens":4096,"output_tokens":180,"tenant_id":"tenant-b","priority_class":"standard","prefix_id":"doc-17","request_type":"rag_qa"}

Replay

PYTHONPATH=src python3 -m yieldos.cli replay --trace path/to/trace.csv --out runs/replay_demo

BurstGPT Adapter

YieldOS-Lite includes a BurstGPT normalizer for the public HPMLL/BurstGPT trace. BurstGPT provides request submission timestamps, request-token counts, response-token counts, model names, and log type. YieldOS maps those fields into the replay schema and writes the normalized sample under the run directory.

PYTHONPATH=src python3 -m yieldos.cli burstgpt-replay \
  --source https://raw.githubusercontent.com/HPMLL/BurstGPT/main/data/BurstGPT_1.csv \
  --limit 800 \
  --time-scale 20 \
  --out runs/burstgpt_sample

--time-scale compresses interarrival times. For example, 20 means 20x higher request rate than the raw timestamps. This keeps the replay explicit: the source trace is external, while the stress level is an experimental knob.