,

Performance Characteristics of Common Network Fabrics

Ethernet

Performance of ethernet networks varies widely. Factors include the switch and NIC manufacturer, firmware settings and system/software settings. Physical layer even plays a role: 10 GigE over RJ45 (10GBase-T) has a higher latency than SFP+ Direct-Attach copper.

Contact our experts to determine a configuration which meets your requirements.

Data RateTheoretical Bandwidth (unidirectional)End-to-End LatencyTechnology
Gigabit Ethernet125 MB/s25 ~ 65 microseconds 
10G Ethernet1.25 GB/s1.3 microseconds (RDMA application) 4 microseconds (sockets application)Mellanox ConnectX-3 VPI
40G Ethernet5 GB/s1.3 microseconds (RDMA application) 4 microseconds (sockets application)Mellanox ConnectX-3 VPI

InfiniBand and Omni-Path Fabrics

Although these fabrics typically offer the highest throughput and lowest latency, much depends on the configuration of the fabric and the method in which your software application accesses the fabric. The figures below describe the best possible performance – contact one of our experts to learn more.

The MPI Bandwidths are measured with large messages and MVAPICH2 MPI. The end-to-end latencies are measured with small messages and presume a single switch connecting two host adapters. Each additional switch hop will add latency (see table below).

Data RateMPI Bandwidth (unidirectional)End-to-End LatencyGeneration
10Gb/s SDR1 GB/s2.6 microsecondsMellanox InfiniHost III
20Gb/s DDR2 GB/s2.6 microsecondsMellanox InfiniHost III
40Gb/s QDR4 GB/s1.07 microsecondsMellanox ConnectX-3
40Gb/s FDR-105.16 GB/s1.07 microsecondsMellanox ConnectX-3
56Gb/s FDR6.82 GB/s1.07 microsecondsMellanox ConnectX-3
100Gb/s EDR12.08 GB/s1.01 microsecondsMellanox ConnectX-4
100Gb/s Omni-Path12.36 GB/s1.04 microsecondsIntel 100G Omni-Path

Larger fabrics require that multiple switches be connected to provide service to all nodes. In such a fabric, each additional switch hop adds a small amount of latency.

Data RateHop LatencyGeneration
40Gb/s QDR0.10 microsecondsMellanox InfiniScale IV
56Gb/s FDR0.20 microsecondsMellanox SwitchX-2
100Gb/s EDR0.09 microsecondsMellanox Switch-IB
100Gb/s Omni-Path0.10 microsecondsIntel 100G Omni-Path
200Gb/s HDR<0.09 microsecondsMellanox Quantum

See also: Performance Characteristics of Common Transports and Buses

You May Also Like

  • Knowledge Center

    Common Maintenance Tasks (Clusters)

    The following items should be completed to maintain the health of your Linux cluster. For servers and workstations, please see Common Maintenance Tasks (Workstations and Servers). Backup non-replaceable data Remember that RAID is not a replacement for backups. If your system is stolen, hacked or started on fire, your data will be gone forever. Automate this…

  • Knowledge Center

    Detailed Specifications of the “Ice Lake SP” Intel Xeon Processor Scalable Family CPUs

    This article provides in-depth discussion and analysis of the 10nm Intel Xeon Processor Scalable Family (formerly codenamed “Ice Lake-SP” or “Ice Lake Scalable Processor”). These processors replace the previous 14nm “Cascade Lake-SP” microarchitecture and are available for sale as of April 6, 2021. The “Ice Lake SP” CPUs are the 3rd generation of Intel’s Xeon…

  • Knowledge Center

    Detailed Specifications of the AMD EPYC “Milan” CPUs

    This article provides in-depth discussion and analysis of the 7nm AMD EPYC processor (codenamed “Milan” and based on AMD’s Zen3 architecture). EPYC “Milan” processors replace the previous “Rome” processors and are available for sale as of March 15th, 2021. These new CPUs are the third iteration of AMD’s EPYC server processor family. They are compatible…