BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260605T154541Z
LOCATION:Bldg. 8 - Room B 101
DTSTART;TZID=Europe/Stockholm:20260629T170000
DTEND;TZID=Europe/Stockholm:20260629T173000
UID:submissions.pasc-conference.org_PASC26_sess126_msa255@linklings.com
SUMMARY:KVCache Data Management in Distributed GenAI Workloads: Characteri
 stics, Requirements, and Challenges
DESCRIPTION:Animesh Trivedi (IBM Research)\n\nGenerative AI (GenAI) worklo
 ads — code completion, intelligent agents, semantic RAG, and recommender s
 ystems — have become widely adopted across many domain-specialized setting
 s. These systems run on large-scale, high-performance infrastructure of th
 ousands of GPUs, storage, and networking devices working together to produ
 ce high-quality responses. While response generation is compute-intensive,
  the data and state management side of GenAI is drawing growing attention.
  One particularly important piece of inference-time state is the Key-Value
  (KV) tensor. KV tensors are generated once but can be reused many times u
 nder the right sharing conditions, can be very large (hundreds of gigabyte
 s to terabytes), and significantly affect the operational efficiency of di
 stributed inference platforms — shaping cost per token, energy use, and re
 quests per second.\nIn this talk, I will introduce KV caches and the prope
 rties of this emerging data type: how they are generated, stored, and shar
 ed today in state-of-the-practice open-source systems such as vLLM and llm
 -d. I will then outline the core data management challenges and opportunit
 ies they present, and describe how we are exploring them through real syst
 em building and simulation.\n\nDomain: Climate, Weather, and Earth Science
 s, Life Sciences, Physics, Computational Methods and Applied Mathematics\n
 \nSession Chair: Vincenzo Eduardo Padulano (CERN)\n\n
END:VEVENT
END:VCALENDAR