There’s some context needed for this, too. So the workload that was reported being halved was a stress test with 1024 clients hammering requests to an AWS VM with a 96-core Graviton4 CPU - so 96 threads for Postgres - and 120GB shared_buffers. The disk were 12 SSDs with 32k IOPS each.

In this very specific and artificial benchmark, this resulted in a -49% performance drop. The thing is, for this to become a problem, a huge amount of page faults have to occur disrupting the lock holders, leading to lock contention. So this requires a machine with a massive amount of clients attempting requests at the same time. Then there will be a large amount of page faults when using the standard 4 kilobyte pages. However, in such a setup you would gain a sizeable amount of performance with huge pages anyway and you will not see anything else running on a machine with those specs, as it would be silly to install something else on a server handling that many clients with those kinds resources.

If you have a known amount of memory your database will receive, then allocating almost all memory as huge pages was always the reasonable thing to do, as not using huge pages meant leaving a significant amount of performance on the road.

Here is an article about using huge pages, and they have a test at the end where with just 180 clients they see a 9-20% performance boost. Note that the article is not about the kernel 7.0 regression, just tuning Postgres in older setups. Huge pages won’t lead to more than 1-5% speedup with smaller setups but you should still do it for a few reasons like memory fragmentation.

Either way, for a large scale setup, like the one the AWS engineer set up, not using huge pages would be absolutely stupid, and that has nothing to do with kernel 7.0.

tldr: performance regression only happens in a synthetic stress test for a huuuge database instance that has not to be configured correctly as well. Plus, Postgres can use RSEQ time slices to recover the performance even in this benchmark if they so desire. That is why this was considered not breaking user space.

This is a massive nothing burger.