The increasing demand for specialized hardware acceleration across a myriad of modern applications necessitates innovative architectures that seamlessly integrate these capabilities into the Cloud-Edge continuum. This paper introduces RECS, a modular and scalable microserver platform specifically designed to facilitate heterogeneous hardware acceleration within distributed Cloud-Edge environments. By integrating a diverse range of microserver architectures—including CPUs, GPUs, FPGAs, and specialized accelerators—RECS enables the configuration of optimized processing platforms tailored to performance-critical workloads. The platform’s architecture supports orchestration and deployment strategies essential for accelerated workflows, while its robust communication infrastructure ensures efficient data management and fault tolerance across the continuum. The management subsystems within RECS provide sophisticated tools for monitoring accelerated applications, contributing to reliability and security in distributed environments. The paper not only presents the architecture of the RECS platform and the various microservers but also evaluates their performance based on Yolov4. Additionally, it summarizes the performance and efficiency gains achieved for different applications deployed on the platform, primarily within the framework of collaborative research projects. On average, a 3.9x improvement in application performance, as well as a 5.2x improvement in energy efficiency compared to the respective application baselines.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

RECS: A Scalable Platform for Heterogeneous AI Acceleration in the Cloud-Edge Continuum

  • René Griessl,
  • Florian Porrmann,
  • Kevin Mika,
  • Lennart Tigges,
  • Jens Hagemeyer

摘要

The increasing demand for specialized hardware acceleration across a myriad of modern applications necessitates innovative architectures that seamlessly integrate these capabilities into the Cloud-Edge continuum. This paper introduces RECS, a modular and scalable microserver platform specifically designed to facilitate heterogeneous hardware acceleration within distributed Cloud-Edge environments. By integrating a diverse range of microserver architectures—including CPUs, GPUs, FPGAs, and specialized accelerators—RECS enables the configuration of optimized processing platforms tailored to performance-critical workloads. The platform’s architecture supports orchestration and deployment strategies essential for accelerated workflows, while its robust communication infrastructure ensures efficient data management and fault tolerance across the continuum. The management subsystems within RECS provide sophisticated tools for monitoring accelerated applications, contributing to reliability and security in distributed environments. The paper not only presents the architecture of the RECS platform and the various microservers but also evaluates their performance based on Yolov4. Additionally, it summarizes the performance and efficiency gains achieved for different applications deployed on the platform, primarily within the framework of collaborative research projects. On average, a 3.9x improvement in application performance, as well as a 5.2x improvement in energy efficiency compared to the respective application baselines.