Dynamic Data Structures and Memory Management on GPUs
summary
Library demonstrates an on-GPU memory manager that performs variable-size allocations and frees without host intervention and backs resizeable, shrinkable data structures. It implements dynamic hash/table and array primitives that let graph algorithms operate on streaming graph analytics directly on device memory, and exposes kernels for sparse tensor contractions used in machine learning and quantum chemistry. Results show usable end-to-end tools rather than a prototype, with allocator latency and scalability prioritized for high-throughput workloads.
AI-written summary · about
chapters (3)
- 0:00Dynamic data structures on GPUs
- 1:10High-performance GPU data goals
- 2:16GPU memory management without host
related talks
01h 03mEmbedded: Customizing Dynamic Memory Management in C++
01h 00m"Designing (New) C++ Hardware”
20mGray-in-Young: A Generational Garbage Collection for Processing-in-Memory
14mFreedom from memory interference
11mUnderstanding and Tackling the Hidden Memory Latency for Edge-based Heterogeneous
25mBridging the Memory Hotness Gap in Edge Systems with Hotness-Segregated Object Allocation
52mKeynote: Domain-Specific Memory Management?
01h 11mMachine Learning Acceleration in an Embedded World