Hey everyone! I'm currently dealing with some io1 volumes, each capable of 20000 IOPS, while using an r5.16xlarge instance. Despite that, we seem to be hitting a ceiling of 4000 IOPS, which translates to about 530MB/s for our workload. The official documentation states that the r5.16xlarge should provide a baseline of 1700MB/s for a 128kb block size, which we've usually seen. However, today it's been quite poor, and our graphs clearly show we're capped at the 4k limit. Has anyone encountered something similar? Is there a known issue or limitation in the eu-central-1 region that could explain this?
1 Answer
According to the docs, this performance limitation sounds about right. io1 volumes can reach up to 32,000 IOPS but they support a max I/O size of 256 KiB, yielding about 500 MiB/s. So, if you're seeing 4000 IOPS with 128 KiB blocks, that's aligning with expected behavior since throughput caps at that point.
Yeah, that makes sense but it's wild seeing such a drop! I used to work with io2 since it allows specifying throughput, but it seems like some folks are hesitant to move forward. And honestly, gp3 seems to work better for most scenarios.