Before now, we have shared a few posts that explore benchmarking and the analysis of the r3 instance family, as well as other instances. You can find any number of those posts here.
In today’s post, we will venture into the world of network performance, benchmarking the r3 instance family. (Remember, as previously stated in subsequent posts, r3 is the new generation memory optimized instance family.)
For the purposes of this analysis, we followed the same methodology as we did when we recently benchmarked the c3 instance family for network performance. You can read about those results here.
So, let’s get started.
We basically ran three types of tests:
- Iperf test → to measure bandwidth. (You can get more information about the Iperf tool from four posts published earlier this year exploring the core of AWS instances: here, here, here, and here.)
- Sockperf ping-pong test → to measure network latency on tcp. (Sockperf is a network performance measuring tool. You can get more details on sockperf here. In ping-pong mode, sockperf can be used to measure network latency.)
- Ping test → to measure network latency using ICMP ping.
All of these benchmarks were run between two identical instances in the EC2 classic network, as well as within a VPC. Below is a brief summary in table format of all the benchmark results. (Note that r3 instance types support enhanced networking when within a VPC.)
As the table shows, we see a slight improvement in network bandwidth when VPC is used. We also see that for r3.8xlarge the bandwidth we measured within a VPC is at nearly 10 Gbps.
Here is a bar graph showing the network performance per hourly rent paid.
In the next table, we outline the results for network latency measurement using sockperf.
There’s a significant reduction in latency when VPC is used. But, otherwise, there are no major differences in latency between different instance types.
Finally, let’s take a look at the results we got when we used ping to measure latency. (Note that we used the default options of ping.)
As we discovered during the sockperf latency, even here, with ping, we see a major reduction in latency within a VPC network. But, in this case, the latencies of different instance types are quite similar.
With these benchmarking results, as always, our aim is to help you make the right choices when working with AWS instance families and the various tools used to measure them.
We’d like to keep the conversation going until the next benchmarking post. Let us hear about some of your findings and thoughts in regard to an analysis of network performance on the r3 instance family, or any other instances. Send us your inquiries at email@example.com. Or, get more information about AWS instances' benchmarking by visiting us at flux7.com/blogs/benchmarks.