It’s important to have the correct network switch with Emulex OneConnect® 10Gb Ethernet (10GbE) Network Xceleration™ (NX) OCe12000-D adapters for low latency applications.
SC12 is in full swing this week, and I am sure we will be hearing all about the latest technology for High Frequency Trading (HFT), High Performance Computing (HPC), RDMA, low latency, InfiniBand (IB), 40GbE and a few others. In line with the low latency, Emulex has been testing the impact of having the correct network switch in HFT environments. In many cases, a network switch is just a network switch with a few bells and whistles. However, when it comes to having the correct network infrastructure to achieve a few more thousand trades than the competition, choosing the right switch can make a difference.
Emulex Technical Marketing engineers (TMEs) tested Emulex’s OCe12000-D low latency adapters on two servers connected back-to-back with no switch in between. Both User Datagram Protocol (UDP) and Transmission Control Protocol (TCP) were tested with and without transparent acceleration (TA) to show the difference between the two protocols. As expected, UDP should have lower latency when compared to TCP, but would a switch make a difference? UDP is small and lightweight so there’s no error correction, which is why it is so quick. There is some error-checking with UDP, but there’s no recovery option. Because UDP is a small lightweight protocol and sits on top of IP, there’s no ordering of messages and no tracking of connections. TCP, on the other hand, has to set up a connection before any data can be sent, and it also does check for reliability and congestion control adding to the overhead. TCP can guarantee the message arrives intact as it was meant to be sent. On the other hand, UDP has no guarantee the messages sent will arrive.
Emulex TMEs tested two switches, one from Gnodal GS4008, and the other a standard 10GbE network switch. Each server had a dual-port OCe12000-D adapter. One port was connected to the Gnodal switch and the second port was connected to the 10GbE switch. The same was done on the second server. We ran a simple utility to test for both UDP and TCP messages. The utility used is called tcp_pinpong which is a simple point-to-point test, showing basic functionality between one OCe12000-D sender adapter port to another OCe12000-D receiver adapter port. There was minimal port switch configuration on both switches, so these results will vary when the switches are tuned to vendor switch recommendation.
The latency test results for both of the switches were compared to the server in back-to-back tests. The servers connected back-to-back were as low as 2.8µs for UDP and 3.8µs for TCP. Just like the server back-to-back test, we used Emulex’s FastStack ™ DBL to demonstrate TA by running dblrun command in front of the tcp_pingpong command. The image below show results of the Emulex OCe12000-D adapter when connected back-to-back with a Gnodal low latency switch and a regular 10GbE network switch. The importance of having the right configuration in your infrastructure can result in more trade transactions versus a standard 10GbE switch.
- Speedometer 1: Highlights the lowest latency simulation scenario utilizing a connection of back-to-back servers with no switch in between
- Speedometer 2: Highlights low latency numbers utilizing a Gnodal 10GbE low latency switch in between the servers
- Speedometer 3: Highlights: latency numbers utilizing regular 10GbE switch in between the servers.
We hope to see you at SC12 this week at our booth, #632, to see low latency demonstrations for the HFT market!