Server with 2xSSD in mdadm mirror slows down one of the disks. How to find the reason?

Question

linux

Colleagues, faced the following problem:
There is a server in hetzner. The server was under a relatively low load, everything worked fine. The disk subsystem is 2 ssd drives in a soft mirror. After some time, the monitoring reported about problems in the smart of one of the disks, about which I created a support request and the disk was changed. Mirrored everything back, the server worked fine. But it was all at night, and when the day came and the usual load went on, the services (sites) began to slow down wildly. I looked in atop and saw something terrible there:

sdb is the same "new" disk.
Of course, the disk was delivered not fresh and heavily used, but they even agreed to a second replacement, the situation repeated itself. In this case, the synchronization of disks takes place at 200MB per second.

Answer 1

It is quite likely that TRIM does not work, SSDs are ordinary user ones, over provisioning is not allocated.
As a result, we get - during synchronization, a large amount of recording, the reserve of clean cells ends, and quite naturally, wild brakes begin on the new disk, but since the mirror does not work faster than the slowest of the disks of its components, it slows down the entire array.