Now 6916 visitors
Today:0 Yesterday:0
Total: 153664 212S 50P 4R
2026-10-20, Week 43
Member Login
TACT Journal Homepage
Call for Paper
Committee
Find My Paper
Author Homepage
Statistics & History
Welcome Message
Paper Procedure
FAQ
Paper Archives
Outstanding Papers
Program & Proceedings
Presentation Platform
Hotel & Travel Info
Photo Gallery
Scheduler Login
Seminar
Archives Login
Sponsors


















Virual Oral Question and Answer
IEEE/ICACT20260329 Question.8
Questioner: wangsy@dsp.ac.cn    2026-02-08 오후 2:27:59 		IEEE/ICACT20260329 Answer.8
Answer by Auhor yangyn@dsp.ac.cn   2026-02-08 오후 2:27:59
What is your optimization goal? (Minimum iteration time? Also including partitioning + mapping?) The optimization goal in our paper is explicitly to minimize the per-iteration training time (Titer​), as defined in Section III (Problem Definition and System Modeling). This objective encompasses co-optimizing two critical aspects: (1) partitioning the model into contiguous stages, and (2) mapping these stages to data centers with specific GPU replica counts. The formulation Titer​=(M+S−1)×maxs​Tstages​integrates pipeline parallelism costs, where Tstages​depends on stage computation/communication times influenced by partitioning and mapping decisions. Thus, the goal holistically addresses partitioning and mapping to reduce latency while adhering to privacy constraints, ensuring efficient multi-DC training.

Select Voice