Ever since Amazon announced AWS EC2 C5D instances, we at TotalCloud.io — as AWS practitioners — have been digging deep into the technicalities of it. We found some interesting stuff, and we’d like to share our findings for the greater good.
For those unaware, AWS EC2 C5D instances are high-performance block storage instances. Introduced sometime during May 2018, they are said to be ideal for applications that need access to high-speed, low latency local storage. This has made it media vertical’s favorite instance out of the lot. If you want to know more about its applications and other details, click here.
To test the performance of C5D instances, we launched a C5D.large instance in US-east 1 region with 9GB Amazon Elastic Block Store + 50GB Amazon EC2 NVMe Instance Storage.
In the same region and availability zone, we launched a C5.large instance with 54 GB EBS volume and T2.small instance for comparison with no NACLs in the way, and open security groups between the instances. All instances were running Amazon Linux 2 OS (ami-0a5e707736615003c), and patched up to date as of October 2018.
Before we walk you through the benchmarking results, let’s have a look at the pricing comparison between C5D.Large, C5.Large and T2.Small on-demand instances:
We ran the Sysbench tool to calculate all prime numbers up to 20,000 to compare and contrast the pure compute performance of C5D instances with C5 instance as well as the the most popular instance type, T2.
#1 | The CPUs all come out at the same score, as all three instances have CPUs of the same capability.
#2 | The C5D vastly outperforms the C5 and T2 in File IO due to the nvme disk, with nearly 4 times the read and write capability of a C5 instance. AWS uses the same underlying disk technology for C5 and T2. Hence the results for C5 and T2 are similar.
#3 | The disk latency measure (IOping) shows that the nvme disk has half the latency of the C5. This is a dramatic difference, and has significant implications for latency-sensitive workloads such as real-time analytics.
#4 | The network measures are essentially the same for C5D and C5 instances, as expected, but significantly improved over the T2 instance.
#5 | The compile time is interesting. The task hits a CPU bottleneck at this point. Although the C5D is faster every time, it’s hard to find a repeatable real-world task that really stretches the capability of the new nvme disk.
#6 | The test showed that the CPU maxes out to 100% to compile Linux. This indicates that C5D is a little faster. Even though compiling Linux C5D is a very disk intensive operation, the CPU was the ultimate bottleneck of the task in our test case.
Just to add to these tests, the C5 instance took 8.10 seconds to unpack the Linux kernel. The C5D took 8.06 seconds to unpack.
There were several tests conducted over a week to benchmark C5D.Large instance’s performance. Hope these test results gives you the real perspective of the new C5D instances.
If you’d like to know more about these test results, write to us at [email protected] or tweet to us at @totalcloudio. Also, do share your views on this post. We would love to hear your thoughts.
Originally published at blog.totalcloud.io on November 9, 2018.