Competitive Benchmark Emulex OCe11102-N and Mellanox ConnectX-3 EN on Windows Server 2008 R2 Performance Tests Simulating Real-World Workloads
Table of contents At a Glance... 3 Introduction... 3 Performance... 5 Test Environment... 5 Throughput Results - Summary... 8 Results - CPU Utilization Effectiveness... 8 Conclusions... 11
At a Glance Data centers have many choices for 10Gb Ethernet (10GbE) server adapters. This competitive benchmark compares the Emulex OneConnect OCe11102 10GbE adapter to the Mellanox ConnectX -3 EN adapter. Emulex Advantages: Near full line rate throughput with large messages (1KB and higher) achieving 24 percent higher throughput for unidirectional data flow with 1500 byte MTU frames Throughput 52 percent higher than the Mellanox adapter for bidirectional data flow with 9000 byte Jumbo MTU packets (1KB and higher messages) CPU effectiveness 42 percent higher than Mellanox for 9000 byte Jumbo MTU frames Option to enable iscsi or FCoE storage networking connectivity with full protocol offload Multi-protocol management Introduction ESG estimates that 67 percent of surveyed data centers have either implemented 10GbE as their server-to-server interconnect or actively plan to move from 1GbE to 10GbE. The ramp up of 10GbE on new servers is the confluence of multiple driving factors: Explosive growth of Web 2.0 and customer facing mobile applications I/O aggregation resulting in increasing virtualization densities (virtual workloads per physical host), driven by the launch of more powerful Intel processors (Intel s Xeon E5 [Romley]) Cloud-driven demand for isolated tenant networks with future proofed scalability Reducing per-port cost vis-à-vis legacy 1GbE This growth of 10GbE networking, coupled with Microsoft Windows Server s dominant server market share of more than 45 percent 1 underscores the importance of reviewing the performance of Emulex s OneConnect OCe11102-N 10GbE adapter in this computing configuration. Emulex is the 10GbE worldwide port market share leader 2. This competitive benchmark summarizes a performance study conducted by Emulex Labs comparing the NIC performance of the OCe11102-N with the Mellanox ConnectX-3 EN using Windows Server 2008 R2. 1 IDC Worldwide Quarterly Server Tracker 4Q 11 2 Crehan Research Calendar Q1-2012 report released May 2012 3 Emulex Competitive Benchmark OCe11102-N and ConnectX-3 EN on Windows Server 2008 R2
Emulex OneConnect Universal CNA Platform The Emulex OneConnect OCe11102 Universal Converged Network Adapter (UCNA) is a thirdgeneration, high performance 10GbE I/O platform that provides server connectivity for network and storage traffic with multi-function server adapters. With OneConnect, Emulex is offering a compelling choice that provides key advantages that include: Available for both rack-mount and blade servers Compatible with LAN on Motherboard (LOM) adapters based on the OneConnect platform that are bundled with new servers from leading system vendors Performance ready with available TCP/IP stateless and TCP Chimney offload Comprehensive and unmatched OS support Network convergence enabled with support for IEEE Data Center Bridging (DCB) VMQ Optimization to support virtualization performance on Windows Server 2008 R2 Universal Multi-Channel (UMC) capability that allows multiple PCI functions to be created on each adapter port 4 Emulex Competitive Benchmark OCe11102-N and ConnectX-3 EN on Windows Server 2008 R2
Option to enable support for iscsi or FCoE with protocol offload when and where needed Multi-protocol management throughout the data center with OneCommand Manager Please visit http://www.emulex.com/artifacts/9f54ab0a-f6fc-4baa-a154-95282c00afe5/elx_ds_all_oce11102-n.pdf for a complete discussion of the OCe11102 platform s capabilities. Performance Performance is another key issue to be considered when evaluating 10GbE adapters. To help answer the performance question, Emulex Labs conducted a series of tests on Windows Server 2008 R2 performance for the OneConnect OCe11102-N and the Mellanox ConnectX-3 EN. A key area of focus for this test effort was a simulation of a real-life production usage environment for a better measure of performance achieved by opening multiple simultaneous TCP/IP connections during the testing. Many vendors do not pursue this methodology, thus presenting a more optimistic performance profile than is realizable in a data center environment. Test Environment Dual-Processor Six-Core Intel Xeon 5680 based Servers Windows Server 2008 R2 Enterprise Edition Windows TCP Global Settings o RSS = Enabled o Receive Window Auto tuning Level = Normal o Add-on congestion Control Provider = cctn o ECN Capability = Disabled o IPv6, QoS Packet Scheduler, File & Printer Sharing = Disabled o Link-Layer Topology Discovery Mapper IO Driver = Enabled o Link-Layer Topology Discovery Responder = Enabled Number of simultaneous TCP Connections = 24 (1 Port), 48 (2 Ports) TCP/IP and Chimney Offload capabilities turned off Max power to all CPUs - No balance Power Savings, Hyperthreading = Enabled IXIA IxChariot Version 7.19 EA Test Tool OCe11102-N Driver Version 4.0.346.0, Firmware version 4.0.360.0 Mellanox ConnectX-3 Driver version MNLX_WinOF_VPI-3.0.0, Firmware version 2.7.2 5 Emulex Competitive Benchmark OCe11102-N and ConnectX-3 EN on Windows Server 2008 R2
Results - Throughput Testing Performance was tested for both 1500 byte and 9000 byte ( Jumbo ) Ethernet MTU frame sizes. Representative data is shown in Figures 1 and 2. For larger message sizes (1K Byte and higher), with 24 simultaneous connections, the OCe11102-N achieves almost full line rate throughput, with peak throughput higher by 127percent or 5.82Gbps, than the ConnectX-3 shown in Figure 1. Unidirectional Throughput Standard Frames Figure 1 6 Emulex Competitive Benchmark OCe11102-N and ConnectX-3 EN on Windows Server 2008 R2
At the other boundary condition, with both ports transmitting and receiving 9000 byte MTU Jumbo frames, while supporting 96 simultaneous TCP connections, the OCe11102-N s peak throughput is higher by 74 percent or 5.82Gbps, than the ConnectX-3 shown in Figure 2. Bidirectional Throughput Jumbo Frames Figure 2 7 Emulex Competitive Benchmark OCe11102-N and ConnectX-3 EN on Windows Server 2008 R2
Throughput Results Summary Figure 3 Peak Throughput Difference 1500 Byte MTU Gbps 9000 Byte MTU Gbps Emulex MLNX Emulex MLNX One Port Unidirectional +5.2 +.36 One Port Bidirectional +4.4 +4.2 Two Port Unidirectional +11.2 +6.0 Two Port Bidirectional +10.9 +16.8 Throughput results summarized in Figure 3 demonstrate the OCe11102-N s performance is superior in a majority of test conditions and virtual parity in one of the other two conditions. Results - CPU Utilization Effectiveness Another measure of network adapter performance is its CPU effectiveness performance. The percentage of CPU utilized to transmit and receive data directly impacts the server CPU s ability to perform other tasks, and its ability to achieve higher virtualization workload density. Representative CPU utilization performance measurements were gathered and translated into CPU effectiveness, defined as Gbps/Percent of CPU utilization. Measurements were taken for both Transmit and Receive performance in back-to-back connected servers. 8 Emulex Competitive Benchmark OCe11102-N and ConnectX-3 EN on Windows Server 2008 R2
Results for 1500 byte MTU frames with Bidirectional data flow are shown in Figure 4. The Emulex adapter has peak effectiveness 0.75 Gbps/% CPU higher on E1 and 0.79 Gbps/% CPU higher on E2. Figure 4 9 Emulex Competitive Benchmark OCe11102-N and ConnectX-3 EN on Windows Server 2008 R2
Results for 9000 byte Jumbo MTU frames with Bidirectional data flow are shown in Figure 5. The Emulex adapter has peak effectiveness 2.77 Gbps/% CPU higher on E1 and 2.72 Gbps/% CPU higher on E2. Figure 5 10 Emulex Competitive Benchmark OCe11102-N and ConnectX-3 EN on Windows Server 2008 R2
Conclusions Testing 10GbE network adapter performance with a large number of TCP connections more accurately simulates workloads in production environments and is a better representation of the adapter s capabilities. The Emulex OCe11102-N delivers excellent performance on both throughput and CPU effectiveness measures, with both ports running concurrently, demonstrating its superiority to the Mellanox ConnectX-3 EN. As data centers deploy more multi-core servers, this is the platform of choice to fully leverage their return on investment. It is no wonder why Emulex is the 10GbEworldwide port market share leader. Some of these products may not be available in the U.S. Please contact your supplier for more information. Copyright 2012 Emulex Corporation. The information contained herein is subject to change without notice. The only warranties for Emulex products and services are set forth in the express warranty statements accompanying such products and services. Emulex shall not be liable for technical or editorial errors or omissions contained herein. OneConnect and OneCommand are registered trademarks of Emulex Corporation.ConnectX-3 is a registered trademark in the U.S. and other countries 11 Emulex Competitive Benchmark OCe11102-N and ConnectX-3 EN on Windows Server 2008 R2
World Headquarters 3333 Susan Street, Costa Mesa, California 92626 +1 714 662 5600 Bangalore, India +91 80 40156789 Beijing, China +86 10 68499547 Dublin, Ireland+35 3 (0)1 652 1700 Munich, Germany +49 (0) 89 97007 177 Paris, France +33 (0) 158 580 022 Tokyo, Japan +81 3 5322 1348 Wokingham, United Kingdom +44 (0) 118 977 2929