Most Amazon Web Services users know that there are many EC2 types. However, fewer people may know the difference between m5d, r5d types and their corresponding m5, r5 series. But after reading this article's title, you might guess that these instance types come with a certain capacity of local storage. In AWS's less down-to-earth terminology, this is called instance store.

As we all know, EBS is not actually a local disk on the server where EC2 resides, but rather allocated from a large distributed storage system and connected to the EC2 server through the network, most likely using protocols like iSCSI to map to EC2's block devices. The benefit is independence from specific servers, allowing EC2 to easily migrate between different servers, or even different data centers and server rooms.

Everything has its pros and cons. As we know, reading and writing large files from disks sometimes consumes significant resources. The main metrics for measuring storage performance are throughput and IOPS. In EBS mode, these two metrics depend not only on the disk itself but are also affected by network bandwidth and fluctuations—and by the worst impact at that.

The so-called instance store is the disk plugged into the server where EC2 resides, so performance is not affected by the network. However, the disk cannot be taken away when EC2 migrates, which is the intrinsic reason why this type of storage is lost when powered off—because the next startup may not run on the original server.

Professional IT personnel can easily understand this principle. Generally, the high-speed private network between EBS and servers means the network may not become a storage performance bottleneck. So in most actual use cases, users don't feel much performance difference between the two. However, I recently tested a program that needed to sequentially read/write and process several terabytes of files multiple times. Running it once took half a day, which was quite painful. And with the same data and same program, execution felt sometimes fast and sometimes slow at different times. This EC2 was in the Tokyo region, and I initially suspected it was affected by the busy/idle status of other tenants in the same AZ.

So I modified the instance type from r5.large to r5d.large, and in the operating system created partitions and file systems on the local disk, then mounted it to the data directory accessed by the program. Running the program again, I found the performance had noticeably improved.

In terms of price, the d series isn't much more expensive than the non-d series. My program's input and output data are both on S3, and the local machine doesn't need to permanently store data—the EC2 shuts down after running the program. In this case, using the d-series instance type not only improves performance and shortens runtime, but also requires no payment for data disks after shutdown. It's a rare choice with double benefits and no downsides—recommended for everyone's reference.

What's the difference between instance store and EBS?

Instance Store (Amazon EC2 instance store) User Guide