Abstract


 
 
 This paper discusses how to decrease and stabilize network latency in a Beowulf system. Having low latency is particularly important to reduce execution time of High Perfor- mance Computing applications. Optimization opportunities are identified and analyzed over the different system components that are integrated in compute nodes, including device drivers, operating system services and kernel parameters.
 This work contributes with a systematic approach to optimize communication latency, provided with a detailed checklist and procedure. Performance impacts are shown through the figures of benchmarks and mpiBLAST as a real-world application. We found that after applying different techniques the default Gigabit Ethernet latency can be reduced from about 50 μs into nearly 20 μs.
 
 

Highlights

  • 1.1 Beowulf ClustersInstead of purchasing an expensive and high-end symmetric multiprocessing (SMP) system, most scientists today choose to interconnect multiple regular-size commodity systems as a means to scale computing performance and gain the ability to resolve bigger problems without requiring heavy investments [1] [2] [3].The key driving factor is cost, out-of-the-box hardware components are used together with open source software to build those systems

  • Clusters are nearly ubiquitous at the Top500 ranking listing most powerful computer systems worldwide, clustered systems represent more than 80% of the list (Figure 1)

  • Latency itself can be measured at different levels, in particular communication latency is a performance metric representing the time it takes for information to flow from one compute node into another

Read more

Summary

Beowulf Clusters

Instead of purchasing an expensive and high-end symmetric multiprocessing (SMP) system, most scientists today choose to interconnect multiple regular-size commodity systems as a means to scale computing performance and gain the ability to resolve bigger problems without requiring heavy investments [1] [2] [3]. The key driving factor is cost, out-of-the-box hardware components are used together with open source software to build those systems. In the specific case of academia, open source software provides the possibility to make software stack modifications, enabling innovation and broadening their adoption. As the cheapest network fabrics are the ones being distributed on-board by system manufacturers, Ethernet is the preferred communication network in Beowulf clusters. At the moment Gigabit Ethernet is included integrated on most hardware

Latency
Related Work
Problem Statement
Intel MPI Benchmarks
Methods
Interrupt Moderation
Interrupt Balancing
Firewall
Kernel Parameters
Optimization Impact
High Performance Linpack
Optimization Procedure
Conclusion
Findings
Future Work
Full Text
Paper version not known

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call

Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.