With the continuous evolution of cloud computing technology, the memory requirements of cloud servers are increasing. Insufficient memory has become the performance bottleneck of many applications. Memory disaggregation has been proposed to solve this problem, which enables computing nodes to utilize the memory of remote memory nodes, thus not only effectively extending the capacity of logical memory, but also significantly improving the resource utilization and elasticity of the whole cluster, providing a more flexible resource allocation scheme for large-scale parallel computing and data processing. However, most of the existing memory disaggregation systems require modification of hardware, operating systems or applications, which are difficult to use. This paper proposes a high-performance memory disaggregation system SwiftSwap, which can directly access remote memory without any modification of hardware, operating system, or application. We design and implement a high performance cache, cache lookup algorithm LDA and cache elimination strategy RM_AwareLRU on the computing node to improve the performance of the system. Through comprehensive evaluation, SwiftSwap has up to 5.1 \(\times \) performance advantage in throughput, 4.8 \(\times \) performance advantage in tail delay, and 6.4 \(\times \) performance advantage against applications running in containers.

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

Fast Memory Disaggregation with SwiftSwap

  • Xiangwei Zhang,
  • Desheng Wang,
  • Weizhe Zhang,
  • Zhiji Yu,
  • Meng Hao

摘要

With the continuous evolution of cloud computing technology, the memory requirements of cloud servers are increasing. Insufficient memory has become the performance bottleneck of many applications. Memory disaggregation has been proposed to solve this problem, which enables computing nodes to utilize the memory of remote memory nodes, thus not only effectively extending the capacity of logical memory, but also significantly improving the resource utilization and elasticity of the whole cluster, providing a more flexible resource allocation scheme for large-scale parallel computing and data processing. However, most of the existing memory disaggregation systems require modification of hardware, operating systems or applications, which are difficult to use. This paper proposes a high-performance memory disaggregation system SwiftSwap, which can directly access remote memory without any modification of hardware, operating system, or application. We design and implement a high performance cache, cache lookup algorithm LDA and cache elimination strategy RM_AwareLRU on the computing node to improve the performance of the system. Through comprehensive evaluation, SwiftSwap has up to 5.1 \(\times \) performance advantage in throughput, 4.8 \(\times \) performance advantage in tail delay, and 6.4 \(\times \) performance advantage against applications running in containers.