Benchmarking: Network Performance Analysis of i instances using Iperf

By Flux7 Labs
March 11, 2014

In a recent post we discussed our methodology for measuring bandwidth with Iperf and how to use Intel’s hardware virtualization drivers to take full advantage of the network card. Iperf is a popular network-benchmarking tool used for measuring bandwidth, packet loss, and delay jitter for the purpose of network tuning, and for creating TCP/UDP data streams. Read our previous post to learn more about Iperf and its installation and setup.

Here we’ll show what we found when using iperf in two different modes. In one mode, we used two instances of the same type within a placement group for both client and server. In the second mode, we put the client and server instances in a VPC and within the same placement group, and then set up “Enhanced Networking mode”. In each mode we ran the client and server on two instances of the same type.

Network Performance When Instances Were Not In a VPC

We ran the benchmark 10 times and discarded the 2 fastest and slowest results. We took the average of the remaining results and used it as the final network bandwidth number for each instance type. Since we ran the tests within Amazon’s network and within the same placement group, the results were fairly consistent. Here are the results for network bandwidth measurement.:

From these results we see that i2.8xlarge and i2.4xlarge share similar network bandwidth. This is because Amazon has a ceiling on bandwidth for i2.8xlarge, which it restricts to about 2.16 Gbits/sec.

Network Performance When Instances Were Within a VPC

When instances ran under VPC with enhanced networking enabled, there was only a slight improvement in network performance compared to the earlier non-VPC setup, except for i2.8xlarge. With enhanced networking enabled, bandwidth of i2.8xlarge was significantly higher, as you can see here:


Here’s the network bandwidth comparison we found when running Iperf both with and without enhanced networking enabled.

An interesting pattern that arises is that, unlike the other characteristics, network performance doesn’t carry strict linear grading with price. The i2.8xlarge instance, by providing access to the complete 10 Gbit ethernet card, was significantly faster than the other instances. If your node is network bound, you’ll definitely get great value for your money. Note also that the i2.xlarge instance also seems less throttled compared to the other instances.