Trading signal latency: dynamic batching breaks 20 ms 99th percentile (P99)
Dynamic batching can boost LLM inference throughput up to 23x, but it can break a 20 ms P99 trading-signal SLO unless queue wait and scheduling are budgeted. Independent technical note mirroring the canonical version: https://hfrtai.com/blog/trading-signal-latency-dynamic-batching-breaks-20-ms-99th-percentile-p99.php
Authors
- Owen Gallagher
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-10-26
- DOI
- https://doi.org/10.5281/zenodo.23682216
- Primary Topic
- Software System Performance and Reliability
- Type
- preprint