NVMe and IRQ Affinity, another problem

[keith.busch@intel.com (Keith Busch)]

On Thu, Apr 05, 2018@12:28:05AM +0000, Young Yu wrote:
> Hello,
> 
> I know that this is another run on the old topic, but I'm still
> wondering what is the right way to bind irq of NVMe-pci devices to the
> cores in local NUMA node.  I'm using kernel 4.16.0-1.el7 on CentOS 7.4
> and the machine have 2 numa nodes as in
> 
> $ lscpu|grep NUMA
> NUMA node(s):          2
> NUMA node0 CPU(s):     0,2,4,6,8,10,12,14,16,18,20,22
> NUMA node1 CPU(s):     1,3,5,7,9,11,13,15,17,19,21,23
> 
> I have 16 NVMe devices, 8 per NUMA node, nvme0 to 7 to the NUMA 0 and
> 8 to 15 to NUMA 1. irqbalance was on by default.  The irq of these
> devices are all bound to the core 0 and 1 regardless of where they are
> physically attached. affinity_hint looks still invalid, however there
> is an effective_affinity that matches with some interrupt
> bounded. cpu_list on mq was pointed to the wrong cores on the NVMe
> devices on NUMA 1. I read it was fixed in kernel 4.3 so not sure
> whether I?m looking at it in a right way.
> 
> Eventually I?d like to know if there is a way to distribute irq of
> each nvme devices to different local cores in NUMA they are attached
> to.

Bad things happened for a lot of servers when the irq spread used
"present" rather than the "online" CPUs, with the "present" CPUs being
oh-so-much larger than what is actually possible.

I'm guessing there's no chance more than 24 CPUs will actually ever
come online in this system, but your platform says 248 may come online,
so we're getting a poor spread for what is actually there.

I believe Ming Lei has an IRQ affinity patch set that may be going in
4.17 that fixes that.

In the meantime, I think if you add kernel paramter "nr_cpus=24",
that should get you a much much better affinity for submission and
completion sides.