We found that NVMe hosts have a really good price to performance ratio for certain tasks, like compiling. For example, an NVMe host might cost just 20% more than an non-NVMe host, but might speed up tasks by more than 2x. This means that using NVMe hosts will both speed up tasks while also making them cheaper.

Today we aren't using NVMe hosts for all tasks that would benefit from doing so because we have constraints on how many NVMe hosts we can have. It's not clear if that constraint is self-imposed (via Evg limits) or AWS-imposed.

As a part of this ticket we should investigate where the limit is coming from and see if we can overcome it. We need to do this for both Windows and Linux NVMe hosts.