Now that is a decade old Westmere architecture, none the less would still have easily handled 200+ concurrent, its the DB hoping between the NUMA nodes, the problems is local RAM to cores are faster, but accessing RAM handing to other cores is much much slower, so if for example first set of RAM (allocated to First CPU) is allocated to DB and O/S, and if O/S needs to do something - it swaps the MariaDB <--- this is where trouble begins.
If you can turn it completely off in BIOS, do that, else see if your BIOS can interleave the memory allocation.
in my.cnf add settings
innodb_numa_interleave=1
additionally, add another SSD (256GB would be enough) and put database there alone
how much memory you've allocated to DB?
using OpCache?