More about HKUST
Adaptive Scheduling and Resource Management for Heterogeneous Multimodal Data Pipelines
PhD Qualifying Examination
Title: "Adaptive Scheduling and Resource Management for Heterogeneous
Multimodal Data Pipelines"
by
Mr. Ding PAN
Abstract:
Modern large language models and multimodal foundation models increasingly
rely on large-scale text, image, audio, video, and document corpora, making
data preparation a major bottleneck in the AI development pipeline. Unlike
traditional stream analytics, multimodal data pipelines combine CPU-bound
preprocessing, accelerator-backed AI inference, highly variable inputs, and
large intermediate artifacts, creating new challenges for scheduling and
resource management. This survey reviews adaptive scheduling for
heterogeneous multimodal data preparation pipelines, with a focus on three
core problems. First, we examine why capacity estimation is unreliable for
asynchronous AI operators, whose observed throughput may be distorted by
continuous batching, input-dependent inference cost, upstream starvation,
downstream backpressure, and transient queue dynamics. Second, we discuss
adaptive configuration tuning under workload regime shifts, where changes in
document length, image resolution, video duration, token length, or modality
mix affect both throughput and accelerator memory usage, making out-of-memory
failures a hard safety constraint. Third, we analyze joint scheduling under
fixed heterogeneous resources, where parallelism, placement, and inference
configuration must be coordinated to avoid shifting bottlenecks or increasing
cross-node data movement. We further review machine-learning-based
performance modeling, Bayesian optimization, and stream-processing
autoscaling as key building blocks. The survey concludes that future
schedulers should move beyond isolated local control toward joint end-to-end
optimization that is workload-aware, memory-safe, bandwidth-aware, and
placement-aware.
Date: Monday, 11 May 2026
Time: 3:00pm - 5:00pm
Venue: Room 3494
Lift 25/26
Committee Members: Dr. Binhang Yuan (Supervisor)
Dr. Wei Wang (Chairperson)
Prof. Ke Yi