This Microsoft page gives a (recently maintained) list of best practices for running Linux under Hyper-V which is the situation I'm facing on a customer site. Easy enough to follow the advice, which is as follows:
Add “numa=off” if the Linux virtual machine has more than 7 virtual processors or more than 30 GB RAM.
Linux virtual machines configured to use more than 7 virtual processors should add numa=off to the GRUB boot.cfg to work around a known issue in the 2.6.x Linux kernels. Linux virtual machines configured to use more than 30 GB RAM should also add numa=off to the GRUB boot.cfg.
The question is, what known issue?
The option itself, as documented on kernel.org says the following:
numa=off Only set up a single NUMA node spanning all memory.
I've seen pages suggesting zone_reclaim_mode can cause issues on NUMA systems, e.g. this cassandra documentation page and an older LinkedIn engineering investigation. Is this the same issue?
Any pointers gratefully accepted. Following the advice, but curious to know the cause.
[link] [comments]