| Paper Abstract and Keywords |
| Presentation |
2026-04-09 12:50
AI load balancer according to SLO Kenshiro Wada, Kenzo Okuda, Hiroki Baba, Kentaro Hayashi, Naoki Kimishima, Tomonori Takeda (NTT) NS2026-5 |
| Abstract |
(in Japanese) |
(See Japanese page) |
| (in English) |
In the operational deployment of large language models (LLMs), inference latency is a critical performance metric.
However, actual requests exhibit strong heterogeneity in input length and concurrent execution counts, making optimisation difficult with a single configuration or simple round-robin distribution.
vLLM incorporates features such as Continuous Batching and Chunked Prefill, enabling the configuration of multiple profiles—latency-oriented, throughput-oriented, or long-text-oriented—through settings.
However, in existing deployments, these are applied statically, failing to fully utilise the performance potential.
This paper proposes an AI-based routing method that dynamically switches execution profiles based on request characteristics: input length (short, medium, long), prefix reuse, and concurrent execution count.
It selects the optimal profile for each characteristic and quantitatively evaluates its effectiveness.
Experimental results show the proposed method improves average and p50 latency by approximately 14%, and p95 latency by approximately 7%, compared to Round-Robin.
This demonstrates that profile selection based on request characteristics is an effective and easily implementable optimisation technique in real-world operations. |
| Keyword |
(in Japanese) |
(See Japanese page) |
| (in English) |
AI / LLM / Proxy / / / / / |
| Reference Info. |
IEICE Tech. Rep., vol. 126, no. 2, NS2026-5, pp. 23-28, April 2026. |
| Paper # |
NS2026-5 |
| Date of Issue |
2026-04-02 (NS) |
| ISSN |
Online edition: ISSN 2432-6380 |
Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
| Download PDF |
NS2026-5 |
| Conference Information |
| Committee |
NS |
| Conference Date |
2026-04-09 - 2026-04-10 |
| Place (in Japanese) |
(See Japanese page) |
| Place (in English) |
Yakushima Island Environmental and Cultural Village Center + Online |
| Topics (in Japanese) |
(See Japanese page) |
| Topics (in English) |
Communication traffic theory, Traffic and quality evaluation, Network performance evaluation, QoS/QoE, Reliability and robustness, Traffic and quality management, AI and machine learning, Network and system operation management, High capacity, low latency, many connections, General |
| Paper Information |
| Registration To |
NS |
| Conference Code |
2026-04-NS |
| Language |
Japanese |
| Title (in Japanese) |
(See Japanese page) |
| Sub Title (in Japanese) |
(See Japanese page) |
| Title (in English) |
AI load balancer according to SLO |
| Sub Title (in English) |
|
| Keyword(1) |
AI |
| Keyword(2) |
LLM |
| Keyword(3) |
Proxy |
| Keyword(4) |
|
| Keyword(5) |
|
| Keyword(6) |
|
| Keyword(7) |
|
| Keyword(8) |
|
| 1st Author's Name |
Kenshiro Wada |
| 1st Author's Affiliation |
NTT (NTT) |
| 2nd Author's Name |
Kenzo Okuda |
| 2nd Author's Affiliation |
NTT (NTT) |
| 3rd Author's Name |
Hiroki Baba |
| 3rd Author's Affiliation |
NTT (NTT) |
| 4th Author's Name |
Kentaro Hayashi |
| 4th Author's Affiliation |
NTT (NTT) |
| 5th Author's Name |
Naoki Kimishima |
| 5th Author's Affiliation |
NTT (NTT) |
| 6th Author's Name |
Tomonori Takeda |
| 6th Author's Affiliation |
NTT (NTT) |
| 7th Author's Name |
|
| 7th Author's Affiliation |
() |
| 8th Author's Name |
|
| 8th Author's Affiliation |
() |
| 9th Author's Name |
|
| 9th Author's Affiliation |
() |
| 10th Author's Name |
|
| 10th Author's Affiliation |
() |
| 11th Author's Name |
|
| 11th Author's Affiliation |
() |
| 12th Author's Name |
|
| 12th Author's Affiliation |
() |
| 13th Author's Name |
|
| 13th Author's Affiliation |
() |
| 14th Author's Name |
|
| 14th Author's Affiliation |
() |
| 15th Author's Name |
|
| 15th Author's Affiliation |
() |
| 16th Author's Name |
|
| 16th Author's Affiliation |
() |
| 17th Author's Name |
|
| 17th Author's Affiliation |
() |
| 18th Author's Name |
|
| 18th Author's Affiliation |
() |
| 19th Author's Name |
|
| 19th Author's Affiliation |
() |
| 20th Author's Name |
|
| 20th Author's Affiliation |
() |
| 21st Author's Name |
|
| 21st Author's Affiliation |
() |
| 22nd Author's Name |
|
| 22nd Author's Affiliation |
() |
| 23rd Author's Name |
|
| 23rd Author's Affiliation |
() |
| 24th Author's Name |
|
| 24th Author's Affiliation |
() |
| 25th Author's Name |
|
| 25th Author's Affiliation |
() |
| 26th Author's Name |
/ / |
| 26th Author's Affiliation |
()
() |
| 27th Author's Name |
/ / |
| 27th Author's Affiliation |
()
() |
| 28th Author's Name |
/ / |
| 28th Author's Affiliation |
()
() |
| 29th Author's Name |
/ / |
| 29th Author's Affiliation |
()
() |
| 30th Author's Name |
/ / |
| 30th Author's Affiliation |
()
() |
| 31st Author's Name |
/ / |
| 31st Author's Affiliation |
()
() |
| 32nd Author's Name |
/ / |
| 32nd Author's Affiliation |
()
() |
| 33rd Author's Name |
/ / |
| 33rd Author's Affiliation |
()
() |
| 34th Author's Name |
/ / |
| 34th Author's Affiliation |
()
() |
| 35th Author's Name |
/ / |
| 35th Author's Affiliation |
()
() |
| 36th Author's Name |
/ / |
| 36th Author's Affiliation |
()
() |
| Speaker |
Author-1 |
| Date Time |
2026-04-09 12:50:00 |
| Presentation Time |
25 minutes |
| Registration for |
NS |
| Paper # |
NS2026-5 |
| Volume (vol) |
vol.126 |
| Number (no) |
no.2 |
| Page |
pp.23-28 |
| #Pages |
6 |
| Date of Issue |
2026-04-02 (NS) |
|