From: Petr Tesarik Date: Tue, 25 Sep 2018 09:23:04 +0200 Subject: Turn off NUMA in the kdump kernel References: bsc#1109784, bsc#1102609 Upstream: merged Git-commit: 4c0e78b9923f0dbba3ad1bacc2336d91f7b520e5 On NUMA systems, the crashkernel reserved region may get allocated on a different node than the low region (memory below 4G). With NUMA optimization turned on, the kernel will prefer node-local pages for all page allocations, which may unnecessarily deplete the DMA and DMA32 regions with data that could be allocated from the Normal zone (which is on a remote node). It makes little sense to address this corner case in the memory management system, because nobody would build a machine with almost all memory on a remote node. So, instead, let's turn off NUMA for kdump and accept the potential performance hit. Signed-off-by: Petr Tesarik --- init/load.sh | 1 + 1 file changed, 1 insertion(+) --- a/init/load.sh +++ b/init/load.sh @@ -73,6 +73,7 @@ function build_kdump_commandline() fi # Use deadline for saving the memory footprint commandline="$commandline elevator=deadline sysrq=yes reset_devices acpi_no_memhotplug cgroup_disable=memory nokaslr" + commandline="$commandline numa=off" commandline="$commandline irqpoll ${nr_cpus}" commandline="$commandline root=kdump rootflags=bind rd.udev.children-max=8" case $(uname -i) in