SYSTEMS
Woven Systems Ethernet Fabric Switch Excels in HPL Benchmark Testing
Results Prove Woven’s Ethernet Fabric Delivers Exceptional Performance and Scalability in High-Performance Computing Applications: Woven Systems today announced that rigorous testing using the High Performance Linpack (HPL) benchmark proves Woven’s 10 Gbps EFX 1000 Ethernet Fabric Switch achieves performance levels equivalent to Double Data Rate (DDR) InfiniBand in most High-Performance Computing (HPC) applications. The extensive tests, which were performed on the Smith cluster at AMD’s Development Center lab in Sunnyvale, Calif., revealed that both systems deliver nearly identical results for Gigaflop performance and scalability, as measured by the HPL efficiency rating. Given Ethernet’s ubiquity, familiarity and ease of use, organizations deploying the Woven Ethernet Fabric Switch will be able to achieve significant reductions in total cost of ownership. Details of the test methodology and results are available in a white paper titled “Comparison of 10 GE with DDR InfiniBand Interconnects using the High Performance Linpack Benchmark.” 
The Linpack benchmark is a numerically intensive test that measures the floating point performance of computers, and serves as the test for the TOP500 list produced biannually by the University of Mannheim, University of Tennessee and NERSC/Lawrence Berkeley National Labs to characterize the performance of the world’s fastest supercomputers. The Woven Systems benchmark testing yielded an HPL efficiency rating greater than 80 percent for both the Ethernet Fabric Switch and DDR InfiniBand. HPL efficiency is used to characterize the performance of HPC clusters, and an 80 percent rating is the industry’s threshold for supporting large-scale “Big Iron” systems, such as those from Cray, NEC and IBM, that have been used to solve the largest compute intensive problems. Both solutions also delivered nearly identical linear performance results, measured in billions of floating point operations (Gigaflops), as the number of nodes increased from two to 36 using systems with eight processor cores.
Woven’s Active Congestion Management is the critical feature that enables the Ethernet Fabric Switch to be used in HPC applications where compute patterns are unpredictable. Active Congestion Management automatically balances the total traffic load throughout the multi-path Ethernet mesh fabric in real-time, overcoming the need to utilize static paths that are prone to congestion. Ethernet also affords the opportunity to consolidate the HPC Applications Interconnect network with the data center’s Storage Transport and System Backbone networks, resulting in additional cost savings.
The EFX 1000 Ethernet Fabric Switch is Woven’s flagship offering for HPC cluster interconnects as well as large-scale enterprise and Internet data centers. A single EFX 1000 switch, configurable with up to 144 10 Gbps Ethernet ports, delivers lossless throughput with a maximum port-to-port latency of 1.5 microseconds. Multiple switches can be deployed in a resilient, multi-path Ethernet Fabric supporting up to 4000 non-blocking edge ports. In such multi-switch configurations, the combination of latency-minimizing Active Congestion Management and cut-through switching results in just 4 microseconds between any two pairs of edge ports.
“The EFX 1000 is already being used successfully in some very large compute clusters, so we were expecting excellent results like these from the Linpack benchmark testing,” said Bert Tanaka, Woven’s founder and CTO. “The test results prove that a 10 Gbps Ethernet Fabric is eminently suitable for the majority of HPC applications, such as LS-Dyna, Fluent, Abacus and many others, enabling operators to get performance comparable to InfiniBand with the added simplicity and flexibility of Ethernet.”
The High Performance Linpack (HPL) benchmark is widely used in the High Performance Computing community and has been used for years to qualify HPC clusters for the TOP500 list. The HPL benchmark measures the performance of a distributed, dense linear algebra solver, and reports both the maximum and theoretical peak floating-point performance for the system being tested. The ratio of these results indicates the system’s efficiency, which provides a reliable indication of how the system should scale as problem size increases.
The Linpack benchmark is a numerically intensive test that measures the floating point performance of computers, and serves as the test for the TOP500 list produced biannually by the University of Mannheim, University of Tennessee and NERSC/Lawrence Berkeley National Labs to characterize the performance of the world’s fastest supercomputers. The Woven Systems benchmark testing yielded an HPL efficiency rating greater than 80 percent for both the Ethernet Fabric Switch and DDR InfiniBand. HPL efficiency is used to characterize the performance of HPC clusters, and an 80 percent rating is the industry’s threshold for supporting large-scale “Big Iron” systems, such as those from Cray, NEC and IBM, that have been used to solve the largest compute intensive problems. Both solutions also delivered nearly identical linear performance results, measured in billions of floating point operations (Gigaflops), as the number of nodes increased from two to 36 using systems with eight processor cores.
Woven’s Active Congestion Management is the critical feature that enables the Ethernet Fabric Switch to be used in HPC applications where compute patterns are unpredictable. Active Congestion Management automatically balances the total traffic load throughout the multi-path Ethernet mesh fabric in real-time, overcoming the need to utilize static paths that are prone to congestion. Ethernet also affords the opportunity to consolidate the HPC Applications Interconnect network with the data center’s Storage Transport and System Backbone networks, resulting in additional cost savings.
The EFX 1000 Ethernet Fabric Switch is Woven’s flagship offering for HPC cluster interconnects as well as large-scale enterprise and Internet data centers. A single EFX 1000 switch, configurable with up to 144 10 Gbps Ethernet ports, delivers lossless throughput with a maximum port-to-port latency of 1.5 microseconds. Multiple switches can be deployed in a resilient, multi-path Ethernet Fabric supporting up to 4000 non-blocking edge ports. In such multi-switch configurations, the combination of latency-minimizing Active Congestion Management and cut-through switching results in just 4 microseconds between any two pairs of edge ports.
“The EFX 1000 is already being used successfully in some very large compute clusters, so we were expecting excellent results like these from the Linpack benchmark testing,” said Bert Tanaka, Woven’s founder and CTO. “The test results prove that a 10 Gbps Ethernet Fabric is eminently suitable for the majority of HPC applications, such as LS-Dyna, Fluent, Abacus and many others, enabling operators to get performance comparable to InfiniBand with the added simplicity and flexibility of Ethernet.”
The High Performance Linpack (HPL) benchmark is widely used in the High Performance Computing community and has been used for years to qualify HPC clusters for the TOP500 list. The HPL benchmark measures the performance of a distributed, dense linear algebra solver, and reports both the maximum and theoretical peak floating-point performance for the system being tested. The ratio of these results indicates the system’s efficiency, which provides a reliable indication of how the system should scale as problem size increases. 