One of the main questions that our customers ask us about running RAC in AWS is whether storage performance will be sufficient. With EBS storage it is possible to provision over 100K IOPS and over 3 GB/s of bandwidth per cluster, which is comparable to a dedicated all-flash storage array. However, some applications can benefit from even higher performance, especially those applications that run complex queries and do large full table scans.
With the recent introduction of the i3 family of EC2 instances it is now possible to use local NVMe SSDs for achieving storage performance levels much higher than what all-flash arrays can provide. Each NVMe SSD is comparable to an entire all-flash array. Each i3 instance can have up to 8 of NVMe SSDs. In a 2-node RAC cluster with i3.16xlarge instances we have measured 27 GB/s of read bandwidth.
But the NVMe SSDs in EC2 are ephemeral, which means that data is not retained on them after an instance is powered down or fails over. FlashGrid Storage Fabric provides mechanisms that turn the ephemeral SSDs into a persistent HA storage by mirroring data across multiple instances and multiple availability zones. Hybrid configurations with NVMe SSDs backed up by EBS volumes are also possible.
On a side note, the process of provisioning a RAC cluster in AWS is fully automated now using CloudFormation templates. Just create a CloudFormation stack with a few mouse clicks and take a lunch break. In less than 1.5 hours you will have a fully configured cluster with Grid Infrastructure and Database software installed, patched, and configured. Send us an email if you want to try it.