BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Chicago
X-LIC-LOCATION:America/Chicago
BEGIN:DAYLIGHT
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
TZNAME:CDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
TZNAME:CST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260202T201228Z
LOCATION:Second Floor Atrium
DTSTART;TZID=America/Chicago:20251118T080000
DTEND;TZID=America/Chicago:20251118T170000
UID:submissions.supercomputing.org_SC25_sess537_drs118@linklings.com
SUMMARY:AI-Driven Resource Optimization for High Performance Computing: A 
 Comprehensive Framework
DESCRIPTION:Manikya Swathi Vallabhajosyula (The Ohio State University)\n\n
 Shared HPC centers are often underutilized because jobs are commonly mis-s
 pecified for walltime, memory, and accelerators. This mis-specification ca
 uses queue churn, idle hardware, and long turnaround times. The main chall
 enge is structural: researchers face a steep learning curve across differe
 nt nodes, policies, and cost models. As a result, they often "guess and su
 bmit" with limited guidance. This work introduces the following center-foc
 used solutions that use predictive models to guide scheduling.\n\n(A) Esti
 mators (black-box + white-box): Two complementary predictors estimate runt
 ime and memory usage based on hardware and configuration. Black-box learne
 rs fit from prior runs; white-box models use operator/graph features and s
 caling laws to generalize. When modelled together, they predict resources 
 with limited training data. \n\n(B) HARP framework: HARP systematizes data
  generation, model building, and selection. It selects estimators based on
  measured error under site policy (queue limits, billing), resulting in a 
 policy-compliant plan for walltime, memory, and devices.\n\n(C) Estimator 
 with Scheduler integration: A scheduler composes estimator outputs with TA
 PIS to produce valid submissions, select queues/partitions, and trade off 
 time and cost. Supports resubmission strategies and “what-if” planning.\n\
 n(D) Closed-Loop Orchestration and Path to Agentic Scheduler: Kafka stream
 s job and filesystem signals to the Intelligence Plane, where estimators e
 nforce policies that drive scheduler daemons, data-generation, and orchest
 ration tasks. Future work extends this loop with goal/constraint inference
 , as well as drift-triggered self-updates, enabling autonomous model train
 ing. This is accompanied by an optional LLM for user interaction and decis
 ion explanation, as well as an MCP-ready design for adaptive scheduling an
 d planning.\n\nTag: Research & ACM SRC Posters\n\nRecording: Not Livestrea
 med, Not Recorded\n\nRegistration Category: Technical Program Reg Pass\n\n
 Session Chairs: Kento Sato (RIKEN Center for Computational Science (R-CCS)
 ); Chris Schlipalius (Pawsey Supercomputing Research Centre; Commonwealth 
 Scientific and Industrial Research Organisation (CSIRO), Australia); and A
 nja Gerbes (Georg-August-Universität Göttingen)\n\n
END:VEVENT
END:VCALENDAR