Marc Olson’s Post

Before 2012, EBS had a bit of a performance problem. We ran a highly multi-tenanted block storage fleet on spinning rust and delivered that storage traffic to the EC2 host via a network interface shared between storage and instance traffic. In 2012, we launched provisioned IOPS volumes and the concept of EBS optimized instances, which provide hard separation between instance traffic and storage traffic. This separation is important when these hosts were connected with 10 or even 1 gigabit ethernet. These first EBS optimized instances carved out up to 1,000Mbps network throughput and up to 8,000 IOPS. In some of these early instances, we managed the separation via traffic shaping over a shared interface Over the years we've increased the performance, first by adding dedicated network interfaces, then by offloading EBS onto the early generation of Nitro cards. As Nitro performance, EBS volume performance, and our customer's storage demands have increased, we've also increased the performance available. It's been fun to watch, but today's launch has me pretty stoked--it's 90x more IOPS (720,000) and 150x (150,000 Mbps) more throughput than the first EBS optimized instances. We did this with a combination of new hardware and new software capabilities. A huge congrats go out to all the teams that worked hard on this, and I can't wait to see the next innovation!  A few milestones: 2012-AUG - EBS optimized launches: 1,000 Mbps / 8,000 IOPS 2014-JAN - C4 instances (first nitro offload): 4,000 Mbps / 32,000 IOPS 2016-NOV - R4 instances: 14,000 Mbps / 75,000 IOPS 2017-NOV - C5 instances (first full nitro instance): 19,000 Mbps / 80,000 IOPS 2020-DEC - R5b instances (first storage heavy): 60,000 Mbps / 260,000 IOPS 2022-NOV - C/M/R6in: 100,000 Mbps / 400,000 IOPS 2025-SEP - R8gb: 150,000 Mbps / 720,000 IOPS https://lnkd.in/ge_d7XGg