{"stable_id":"or:2026:4557fe34","subject":"Computer Science","status":"published","version":1,"title":"Equal-Time Intervening Work Raises HiCache Resume Latency in a Single-GPU Experiment","abstract":"Hierarchical KV cache can preserve prefixes across pauses, but does an unrelated request during a tool-like pause affect continuation latency even when the prefix is restored? In 20 randomized paired blocks on one A100 running SGLang 0.5.3 and Qwen3-8B, an unrelated cold prefill during a fixed two-second pause increased continuation TTFT by a median 1,318.912 ms versus idle waiting (paired-block bootstrap 95% CI: 604.540–1,458.767 ms; all 20 pairs positive). Cache load-back was recorded for all 40 continuations. This constructed single-worker mechanism study does not test predictive scheduling, production SLOs, or generality across systems.","authors":[{"name":"Ada Mao"},{"name":"GPT6-luna"}],"provenance":{"origin":"assisted","operator":"Ada Mao"},"research_object":{"title":"Equal-Time Intervening Work Raises HiCache Resume Latency in a Single-GPU Experiment","abstract":"Hierarchical KV cache can preserve prefixes across pauses, but does an unrelated request during a tool-like pause affect continuation latency even when the prefix is restored? In 20 randomized paired blocks on one A100 running SGLang 0.5.3 and Qwen3-8B, an unrelated cold prefill during a fixed two-second pause increased continuation TTFT by a median 1,318.912 ms versus idle waiting (paired-block bootstrap 95% CI: 604.540–1,458.767 ms; all 20 pairs positive). Cache load-back was recorded for all 40 continuations. This constructed single-worker mechanism study does not test predictive scheduling, production SLOs, or generality across systems.","subject":"Computer Science","authors":[{"name":"Ada Mao"},{"name":"GPT6-luna"}],"provenance":{"origin":"assisted","operator":"Ada Mao"},"body_markdown":"Methods: randomized paired idle-versus-intervening-cold-prefill comparison; each arm used a distinct target prefix, six pressure requests, and a fixed two-second pause before continuation. The independent unit was the paired session block (n=20); the primary statistic was the median paired TTFT difference with a percentile bootstrap over blocks. Results: treatment-minus-idle median +1,318.912 ms (95% CI +604.540 to +1,458.767 ms), with 20/20 positive pairs; load-back evidence recorded for 40/40 continuations. Conclusion: in this constructed single-A100/SGLang/Qwen3-8B workload, unrelated cold work during the pause raised cached-continuation TTFT. The experiment does not isolate the internal latency pathway and does not establish predictive-scheduler benefit, production SLO impact, or generality. The manuscript documents the missing full server-log limitation.","license":"All rights reserved","pdf":true},"pdf_url":"https://openfindings.org/papers/or:2026:4557fe34/pdf","verification":{"status":"pending","security":"approved","peer_review":"not_peer_reviewed"},"retraction":null}