BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260730T152640Z
LOCATION:
DTSTART;TZID=America/Los_Angeles:20260728T174600
DTEND;TZID=America/Los_Angeles:20260728T174600
UID:dac_DAC 2026_sess317_RESEARCH1351@linklings.com
SUMMARY:eLLM: Elastic Memory Management Framework for Efficient LLM Servin
 g
DESCRIPTION:Yi Xiong (University of Science and Technology in China), Jial
 e Xu (Shanghai Jiao Tong University), Rui Zhang (Ant Group), Cong Guo (Duk
 e University), Zihan Liu (Shanghai Jiao Tong University), Yangjie Zhou (Na
 tional University of Singapore), Weiming Hu (Shanghai Jiao Tong University
 ), Hao Wu (Ant Group), Boyu Li (University of Science and Technology of Ch
 ina), Junping Zhao (Ant Group), Minyi Guo and Jingwen Leng (Shanghai Jiao 
 Tong University), and zongwei zhu and Xuehai Zhou (University of Science a
 nd Technology of China)\n\nGPU memory management is critical for efficient
  Large Language Model (LLM) serving. LLM memory usage primarily comprises 
 weights, activations, and KV caches. While weights are static, activations
  and KV caches exhibit dynamic and unpredictable behavior, posing signific
 ant memory management challenges. Modern LLM serving systems address this 
 through a dual-level approach: activations inherit static tensor abstracti
 ons from deep learning frameworks, while KV caches employ specialized page
 -table virtualization (i.e., PagedAttention). Although this reduces KV cac
 he fragmentation, the fundamental isolation between activation and KV cach
 e management prevents memory sharing across these spaces, leading to subop
 timal utilization and 20\% throughput degradation.\n\nTo address these lim
 itations, we propose eLLM, an elastic memory management framework. The cor
 e components of eLLM include:(1) Virtual Tensor Abstraction: Decouples the
  virtual address space of tensors from physical GPU memory, creating a uni
 fied and flexible memory pool;(2) Elastic Memory Mechanism: Dynamically ad
 justs memory allocation through runtime memory inflation and deflation, an
 d leverages CPU memory as an extensible buffer;(3) Lightweight Scheduling 
 Strategy: Employs Service-Level Objective (SLO)-aware policies to optimize
  memory utilization and effectively balance performance trade-offs under s
 tringent SLO constraints.\nComprehensive evaluations demonstrate that eLLM
  outperforms state-of-the-art systems, achieving up to 2.32$\times$ higher
  throughput.\n\nTrack: EDA7-II. Physical Design and Verification, Design\n
 \nSession Chairs: Christian Pilato (Politecnico di Milano); Mohsen Imani (
 University of California, Irvine); Flavio Ponzina (San Diego State Univers
 ity); Amin Farshidi (Cadence Design Systems, Inc.); Yonggan Fu (Georgia In
 stitute of Technology, NVIDIA); Dimitrios Tsaras (Huawei, Hong Kong Univer
 sity of Science and Technology (HKUST)); Anh Nguyen (Fujitsu Limited); Rui
 yang Qin (Villanova University); Jiliang Zhang (Hunan University); Meng Li
  (Peking University); and Rassul Bairamkulov (Advanced Micro Devices (AMD)
 )\n\n
END:VEVENT
END:VCALENDAR