Back to search

Cornell University

Programming Frameworks for Improving the Productivity and Performance of Manycore Architectures

Abstract

dc:description.abstract

Manycore architectures integrate hundreds of cores on a single chip by using simple cores and simple memory systems usually based on software-managed scratchpad memories (SPMs). Such architectures are notoriously challenging to program, since the programmers need to manually manage all aspects of data movement and synchronization for both correctness and performance. This manycore programmability challenge is one of the key barriers to achieving the promise of manycore architectures. This thesis presents both domain-specific (HB-PyTorch and HB-Arc) and general-purpose (HB-Rubick) programming frameworks to address the SPM manycore architecture programmability challenge and/or improve performance. HB-PyTorch enables domain experts to easily accelerate off-the-shelf tensor workloads. Evaluation on three real-world dense and sparse tensor workloads suggests these workloads can achieve approximately 2-6x performance improvement when scaled to a future 2,000-core manycore system compared to an 18-core out-of-order CPU baseline, while potentially achieving higher area-normalized throughput and improved energy-efficiency compared to GPGPUs. HB-Arc explores the potential of decoupled access/execute (DAE) mechanisms, and proposes two software-only techniques, naive-software DAE and systolic-software DAE, along with a lightweight hardware access accelerator for further performance benefit. However, being domain-specific limits their scope. General purpose dynamic task parallel programming frameworks offer many advantages over domain-specific frameworks, including more flexibility and better load-balancing. Conventional wisdom suggests a work-stealing runtime, which forms the core of most dynamic task parallel programming models, is ill-suited for manycore architectures. However, HB-Rubick demonstrates that such a runtime is not just feasible on manycore architectures with SPMs, but it can also significantly improve the performance of irregular workloads when executing on these architectures. The proposed dynamic task parallel programming framework enhanced with three optimizations for leveraging unused SPM space achieves 1.2-28.5x speedup on workloads that benefit from our techniques, and only induces minimal overhead for workloads that do not. This thesis provides a small yet important step towards closing the performance and productivity gap of SPM manycore architectures.

Degree

thesis:*
Name thesis:degree_name
Ph. D., Computer Science
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Computer Science
Grantor
Cornell University
Year dc:date.issued
2022

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Cheng, Lin
Chair dc:contributor.chair
  • Batten, Christopher
Committee members dc:contributor.committeemember
  • Sampson, Adrian
  • Zhang, Zhiru

Subjects

dc:subject × 6

Rights

dc:rights
Statement dc:rights
  • Attribution-NoDerivatives 4.0 International
Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Bibid
15644160
Dc Identifier Other
ProQuest Submission ID: 13378
ProQuest Publication ID: 29996900

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Cheng, Lin. Programming Frameworks for Improving the Productivity and Performance of Manycore Architectures. Doctor of Philosophy thesis, Cornell University, 2022. https://hdl.handle.net/1813/112906