HPCA 2024 · HPCA 2025

Scalable and Adaptive Memory Management for Multi-GPU Unified Memory Systems

GRIT and OASIS

University of Pittsburgh · NVIDIA · Ghent University

A research line on dynamic page placement for UVM-enabled multi-GPU systems, from fine-grained per-page adaptation to object-aware runtime management.

No single UVM policy fits all pages, objects, and phases GRIT and OASIS dynamically choose between on-touch migration, access-counter migration, and duplication. UVM Policies On-touch migration Access counter Page duplication GRIT Fine-grained page adaptation Fault-Aware Initiator PA-Table + PA-Cache Neighbor Prediction OASIS Object-aware policy Object O-Table

GRIT adapts policies at page granularity; OASIS moves the decision to object granularity to reduce overhead and design complexity.

Overview

Modern multi-GPU systems use unified virtual memory to simplify programming, but UVM page management suffers from NUMA overheads caused by remote accesses, page migrations, and write invalidations. Existing systems often rely on one policy across the workload, even though different pages, objects, and execution phases prefer different policies.

GRIT and OASIS address this problem through adaptive runtime page management. GRIT dynamically learns page-level behavior and selects page placement policies using page faults, page attributes, and neighboring-page similarity. OASIS builds on the same insight but shifts the control point from pages to application objects, using object tracking and an object policy table to learn access patterns with lower metadata and hardware overhead.

GRIT: Fine-Grained Dynamic Page Placement

HPCA 2024Fault-Aware Initiator

GRIT uses local page faults and page protection faults as lightweight indicators that the current placement scheme is no longer appropriate.

Page MetadataPA-Table and PA-Cache

A software PA-Table records page access attributes, while a small hardware PA-Cache reduces extra memory traffic from metadata lookups.

Spatial LocalityNeighboring-Aware Prediction

GRIT leverages the observation that neighboring pages often share similar access behavior and proactively assigns policies to adjacent pages.

OASIS: Object-Aware Page Management

HPCA 2025Object Tracker

OASIS identifies runtime objects created by managed-memory allocation and associates memory accesses with the corresponding object.

Object MetadataO-Table

Instead of tracking thousands of pages, OASIS tracks object-level policy state in a compact table, reducing metadata and lookup overhead.

Phase AwarenessObject Policy Controller

OASIS resets and relearns object policies across explicit or implicit phase changes, allowing the runtime to adapt as object behavior changes.

From GRIT to OASIS

GRIT: adaptive at page granularity

GRIT shows that a single policy cannot match the diversity of multi-GPU memory behavior. It adapts page placement dynamically and improves performance by selecting among on-touch migration, access-counter migration, and duplication during execution.

OASIS: simpler at object granularity

OASIS observes that pages belonging to the same object usually exhibit consistent access patterns within a phase. This allows the runtime to make fewer, more stable decisions with lower storage and hardware overhead.

Evaluation Highlights

60%GRIT vs. on-touch
64%OASIS vs. on-touch
12%OASIS over GRIT

GRIT demonstrates the benefit of dynamic page placement; OASIS preserves the benefit while reducing metadata and design complexity through object-aware management.

Takeaway

GRIT and OASIS show that UVM page management should be dynamic, data-aware, and phase-aware. Together, they form a path from fine-grained page policy learning to a more practical object-level runtime that improves multi-GPU memory efficiency with lower overhead.

Citation

@inproceedings{wang2024grit,
  note = {*The authors contribute equally.},
  title = {GRIT: Enhancing Multi-GPU Performance with Fine-Grained Dynamic Page Placement},
  author = {Wang, Yueqi and Li, Bingyao and Jaleel, Aamer and Yang, Jun and Tang, Xulong},
  booktitle = {2024 IEEE International Symposium on High-Performance Computer Architecture (HPCA)},
  pages = {1080--1094},
  year = {2024},
  organization = {IEEE},
}

@inproceedings{wang2024oasis,
  title = {OASIS: Object-Aware Page Management for Multi-GPU Systems},
  author = {Wang, Yueqi and Li, Bingyao and Ziad, Mohamed Tarek Ibn and Eeckhout, Lieven and Yang, Jun and Jaleel, Aamer and Tang, Xulong},
  booktitle = {2025 IEEE International Symposium on High-Performance Computer Architecture (HPCA)},
  year = {2025},
  organization = {IEEE},
  pages = {1678-1692},
}