| 袁雨馨* **,李庆文* **,史骁*,赵晓芳*.基于服务器无感知计算架构的并行计算通信框架[J].高技术通讯(中文),2025,35(6):590~603 |
| 基于服务器无感知计算架构的并行计算通信框架 |
| A parallel communication framework based on serverless architecture |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 06. 003 |
| 中文关键词: 服务器无感知计算； 消息传递接口并行计算； 服务质量保障； 弹性资源分配 |
| 英文关键词: serverless computing, MPI parallel computing, QoS guarantee, elastic resource allocation |
| 基金项目: |
| 作者 | 单位 | | 袁雨馨* ** | (*中国科学院计算技术研究所北京 100190)
(**中国科学院大学北京 100049) | | 李庆文* ** | (*中国科学院计算技术研究所北京 100190)
(**中国科学院大学北京 100049) | | 史骁* | (*中国科学院计算技术研究所北京 100190)
(**中国科学院大学北京 100049) | | 赵晓芳* | (*中国科学院计算技术研究所北京 100190)
(**中国科学院大学北京 100049) |
|
| 摘要点击次数: 14 |
| 全文下载次数: 14 |
| 中文摘要: |
| 随着云计算技术的发展与成熟，并行计算在云环境中得到了越来越多的实践。服务器无感知计算作为云计算中的一种新型的应用部署与计算方式，允许用户弹性分配资源并实现负载均衡，并提供了更强的可扩展性和更大的灵活性。然而，无状态的特性导致服务器无感知计算框架并不完全适用于传统并行计算，其中通信是一个关键问题。本文提出了一个具有服务质量(quality of service，QoS)保障的通信框架FreeParallel，旨在基于服务器无感知计算中的函数即服务(function as a service，FaaS)平台构建面向并行计算的通信能力。FreeParallel结合了消息传递接口(message passing interface，MPI)并行计算编程模型，有效地保证了通信服务的质量；并采用代理模型来支持并行函数的识别和转换，并以服务形式灵活部署在多个FaaS或虚拟化平台上。此外，本研究还提出了函数间通信流量的QoS管理策略fmClock，在保证传输公平性的前提下，实现基于请求和限制的通信原语级网络资源分配。实验结果表明，点对点通信场景下FreeParallel与虚拟化平台的覆盖网络相比传输性能略有不足，但比当前服务器无感知计算状态共享方案的传输效率有至少89.5%的提升。并且FreeParallel在集合通信场景下表现极佳，比基线方法提升了59.9%~83.1%。同时，带有fmClock策略的FreeParallel能够实现原语级按比例分配策略，避免了不同原语间请求的交叉干扰，案例表明，策略的加入降低了应用25.0%的完成时间。 |
| 英文摘要: |
| With the development and maturation of cloud computing technology, parallel computing has been increasingly practiced in cloud environments. Serverless computing, as a novel deployment and computing approach in cloud computing, allows users to flexibly allocate resources, achieve load balancing, and provides greater scalability and flexibility. However, the stateless nature of serverless computing makes it not fully suitable for traditional parallel computing, where communication is a key issue. This paper proposes FreeParallel, a communication framework with quality of service (QoS) guarantee, aiming to build communication capabilities for parallel computing based on Function as a Service (FaaS) in serverless computing. FreeParallel combines the message passing interface (MPI) parallel computing programming model, effectively ensuring the quality of communication services. It adopts a proxy model to support the identification and transformation of parallel functions, and is flexibly deployed as a service on multiple FaaS or virtualization platforms. Additionally, this research proposes the QoS management strategy fmClock for inter-function communication, which achieves network resource allocation based on request and limitation while ensuring transmission fairness. Experimental results show that in point-to-point communication scenarios, FreeParallel’s transmission performance is slightly inferior to the overlay network of virtualization platforms, but it improves transmission efficiency by at least 89.5% compared to current serverless computing state sharing frameworks. Moreover, FreeParallel performs exceptionally well in collective communication scenarios, with an improvement of 59.9% to 83.1% compared to baseline methods. Meanwhile, FreeParallel with the fmClock strategy can achieve proportional allocation at the communication primitive level, avoiding cross-interference between requests from different primitives. Case studies demonstrate that the inclusion of the strategy reduces the execution time of applications by 25.0%. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
