New – The Next Generation (I3en) of I/O-Optimized EC2 Instances

Amazon’s Customer Obsession leadership principle says:

Leaders start with the customer and work backwards. They work vigorously to earn and keep customer trust. Although leaders pay attention to competitors, they obsess over customers.

Starting from the customer and working backwards means that we do not invent in a vacuum. Instead, we speak directly to our customers (both external and internal), ask detailed questions, and pay attention to what we learn. On the AWS side, we often hear about new use cases that help us to get a better understanding of what our customers are doing with AWS. For example, large-scale EC2 users provide us with another set of interesting data points, often expressed in terms of ratios between dollars, vCPUs, memory size, storage size, and networking throughput.

We launched the I3 instances (Now Available – I3 Instances for Demanding, I/O Intensive Workloads) just about two years ago. Our customers use them to host distributed file systems, relational & NoSQL databases, in-memory caches, key-value stores, data warehouses, and MapReduce clusters. Because our customers are always (in Jeff Bezos’ words) “divinely discontent”, they want I/O-optimized instances with even more power & storage. To be specific, they have asked us for:

  • A lower price per TB of storage
  • Increased storage density to allow consolidation of workloads and scale-up processing
  • A higher ratio of network bandwidth and instance storage to vCPUs

The crucial element here is that our customers were able to express their needs in a detailed and specific manner. Simply asking for something to be better, faster, and cheaper does not help us to make well-informed decisions.

New I3en Instances
Today I am happy to announce the I3en instances. Designed to meet these needs and to do an even better job of addressing the use cases that I mentioned above, these instances are powered by AWS-custom Intel Xeon Scalable (Skylake) processors with 3.1 GHz sustained all-core turbo performance, up to 60 TB of fast NVMe storage, and up to 100 Gbps of network bandwidth. Here are the specs:

Instance NamevCPUs
MemoryLocal Storage
(NVMe SSD)
Random Read IOPS
(4 K Block)
Read Throughput
(128 K Block)
EBS-Optimized BandwidthNetwork Bandwidth
i3en.large216 GiB1 x 1.25 TB42.5 K325 MB/sUp to 3,500 MbpsUp to 25 Gbps
i3en.xlarge432 GiB1 x 2.50 TB85 K650 MB/sUp to 3,500 MbpsUp to 25 Gbps
i3en.2xlarge864 GiB2 x 2.50 TB170 K1.3 GB/sUp to 3,500 MbpsUp to 25 Gbps
i3en.3xlarge1296 GiB1 x 7.5 TB250 K2 GB/sUp to 3,500 MbpsUp to 25 Gbps
i3en.6xlarge24192 GiB2 x 7.5 TB500 K4 GB/s3,500 Mbps25 Gbps
i3en.12xlarge48384 GiB4 x 7.5 TB1 M8 GB/s7,000 Mbps50 Gbps
i3en.24xlarge96768 GiB8 x 7.5 TB2 M16 GB/s14,000 Mbps100 Gbps

In comparison to the I3 instances, the I3en instances offer:

  • A cost per GB of SSD instance storage that is up to 50% lower
  • Storage density (GB per vCPU) that is roughly 2.6x greater
  • Ratio of network bandwidth to vCPUs that is up to 2.7x greater

You will need HVM AMIs with the NVMe 1.0e and ENA drivers. You can also make use of the new Elastic Fabric Adapter (EFA) if you are using the i3en.24xlarge (read my recent post to learn more).

Now Available
You can launch I3en instances today in the US East (N. Virginia), US West (Oregon), and Europe (Ireland) Regions in On-Demand and Spot form. Reserved Instances, Dedicated Instances, and Dedicated Hosts are available.

Jeff;