@inproceedings{huang-etal-2026-signal,
title = "What Signal Does Preemption Need? Benchmarking Preemption-Target Selection for {SLO}-Aware Reasoning {LLM} Serving",
author = "Huang, Qi Yang and
Shah, Alay Dilipbhai and
Ponnusamy, Pragaash and
Le, Hoang Anh Duy and
Joseph, Thomas A. and
Seligman, Adam and
Xu, Zhaozhuo",
editor = "Atanasova, Pepa and
Haffari, Gholamreza and
Wang, Minghan",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-industry.145/",
pages = "2430--2452",
ISBN = "979-8-89176-460-6",
abstract = "When the key-value (KV) cache fills, a serving engine must choose which request to evict. We present PreemptBench, comparing seven preemption-target policies across seven reasoning-model families and two benchmarks. Our service-level objective (SLO) breach rate counts every accepted request, including late completions, post-acceptance failures, and unfinished work; this survivorship correction prevents deferred backlog from being omitted from the metric. In the low-breach regime, where last-in, first-out (LIFO) breach is below 15{\%}, young-target-biased policies form a lower-breach cluster than policies without this bias, including random and largest-KV, consistent with a stable incumbent cohort; graded eviction depth is secondary. Within a pure remaining-decode rule, replacing the online predictor with a privileged empirical statistical length oracle, constructed from prior temperature-zero runs, yields no significant gain; neither significantly outperforms LIFO. It summarizes a prompt-conditioned length distribution, not the current execution{'}s realization, and is not SLO-optimal. Finally, the uncensored single-stream completion-time tail defines a conceptual feasibility floor, but our traces provide only a censored low-load proxy. Post-hoc rescoring across 600{--}1800-second deadlines finds aggregate memory-freeing reversals at tighter SLOs, but none at 1800 seconds. Thus LIFO is a strong conditional baseline, not a universal optimum."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="huang-etal-2026-signal">
<titleInfo>
<title>What Signal Does Preemption Need? Benchmarking Preemption-Target Selection for SLO-Aware Reasoning LLM Serving</title>
</titleInfo>
<name type="personal">
<namePart type="given">Qi</namePart>
<namePart type="given">Yang</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alay</namePart>
<namePart type="given">Dilipbhai</namePart>
<namePart type="family">Shah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pragaash</namePart>
<namePart type="family">Ponnusamy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hoang</namePart>
<namePart type="given">Anh</namePart>
<namePart type="given">Duy</namePart>
<namePart type="family">Le</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thomas</namePart>
<namePart type="given">A</namePart>
<namePart type="family">Joseph</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Adam</namePart>
<namePart type="family">Seligman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhaozhuo</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pepa</namePart>
<namePart type="family">Atanasova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gholamreza</namePart>
<namePart type="family">Haffari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minghan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-460-6</identifier>
</relatedItem>
<abstract>When the key-value (KV) cache fills, a serving engine must choose which request to evict. We present PreemptBench, comparing seven preemption-target policies across seven reasoning-model families and two benchmarks. Our service-level objective (SLO) breach rate counts every accepted request, including late completions, post-acceptance failures, and unfinished work; this survivorship correction prevents deferred backlog from being omitted from the metric. In the low-breach regime, where last-in, first-out (LIFO) breach is below 15%, young-target-biased policies form a lower-breach cluster than policies without this bias, including random and largest-KV, consistent with a stable incumbent cohort; graded eviction depth is secondary. Within a pure remaining-decode rule, replacing the online predictor with a privileged empirical statistical length oracle, constructed from prior temperature-zero runs, yields no significant gain; neither significantly outperforms LIFO. It summarizes a prompt-conditioned length distribution, not the current execution’s realization, and is not SLO-optimal. Finally, the uncensored single-stream completion-time tail defines a conceptual feasibility floor, but our traces provide only a censored low-load proxy. Post-hoc rescoring across 600–1800-second deadlines finds aggregate memory-freeing reversals at tighter SLOs, but none at 1800 seconds. Thus LIFO is a strong conditional baseline, not a universal optimum.</abstract>
<identifier type="citekey">huang-etal-2026-signal</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-industry.145/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>2430</start>
<end>2452</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T What Signal Does Preemption Need? Benchmarking Preemption-Target Selection for SLO-Aware Reasoning LLM Serving
%A Huang, Qi Yang
%A Shah, Alay Dilipbhai
%A Ponnusamy, Pragaash
%A Le, Hoang Anh Duy
%A Joseph, Thomas A.
%A Seligman, Adam
%A Xu, Zhaozhuo
%Y Atanasova, Pepa
%Y Haffari, Gholamreza
%Y Wang, Minghan
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-460-6
%F huang-etal-2026-signal
%X When the key-value (KV) cache fills, a serving engine must choose which request to evict. We present PreemptBench, comparing seven preemption-target policies across seven reasoning-model families and two benchmarks. Our service-level objective (SLO) breach rate counts every accepted request, including late completions, post-acceptance failures, and unfinished work; this survivorship correction prevents deferred backlog from being omitted from the metric. In the low-breach regime, where last-in, first-out (LIFO) breach is below 15%, young-target-biased policies form a lower-breach cluster than policies without this bias, including random and largest-KV, consistent with a stable incumbent cohort; graded eviction depth is secondary. Within a pure remaining-decode rule, replacing the online predictor with a privileged empirical statistical length oracle, constructed from prior temperature-zero runs, yields no significant gain; neither significantly outperforms LIFO. It summarizes a prompt-conditioned length distribution, not the current execution’s realization, and is not SLO-optimal. Finally, the uncensored single-stream completion-time tail defines a conceptual feasibility floor, but our traces provide only a censored low-load proxy. Post-hoc rescoring across 600–1800-second deadlines finds aggregate memory-freeing reversals at tighter SLOs, but none at 1800 seconds. Thus LIFO is a strong conditional baseline, not a universal optimum.
%U https://aclanthology.org/2026.emnlp-industry.145/
%P 2430-2452
Markdown (Informal)
[What Signal Does Preemption Need? Benchmarking Preemption-Target Selection for SLO-Aware Reasoning LLM Serving](https://aclanthology.org/2026.emnlp-industry.145/) (Huang et al., EMNLP 2026)
ACL