[RFC 0/2] genirq: take device NUMA node into account for managed IRQs

[Stefan Hajnoczi]

Devices with a small number of managed IRQs do not benefit from spreading
across all CPUs. Instead they benefit from NUMA node affinity so that IRQs are
handled on the device's NUMA node.

For example, here is a machine with a virtio-blk PCI device on NUMA node 1:

  # lstopo-no-graphics
  Machine (958MB total)
    Package L#0
      NUMANode L#0 (P#0 491MB)
      L3 L#0 (16MB) + L2 L#0 (4096KB) + L1d L#0 (32KB) + L1i L#0 (32KB) + Cor=
e L#0 + PU L#0 (P#0)
    Package L#1
      NUMANode L#1 (P#1 466MB)
      L3 L#1 (16MB) + L2 L#1 (4096KB) + L1d L#1 (32KB) + L1i L#1 (32KB) + Cor=
e L#1 + PU L#1 (P#1)
      HostBridge
        PCIBridge
          PCI c9:00.0 (SCSI)
            Block "vdb"
    HostBridge
      PCIBridge
        PCI 02:00.0 (Ethernet)
          Net "enp2s0"
      PCIBridge
        PCI 05:00.0 (SCSI)
          Block "vda"
      PCI 00:1f.2 (SATA)

Currently the virtio5-req.0 IRQ for the vdb device gets assigned to CPU 0:

  # cat /proc/interrupts
             CPU0       CPU1
  ...
   36:          0          0   PCI-MSI 105381888-edge      virtio5-config
   37:         81          0   PCI-MSI 105381889-edge      virtio5-req.0

If managed IRQ assignment takes the device's NUMA node into account then CPU 1
will be used instead:

  # cat /proc/interrupts
             CPU0       CPU1
  ...
   36:          0          0   PCI-MSI 105381888-edge      virtio5-config
   37:          0         92   PCI-MSI 105381889-edge      virtio5-req.0

The fio benchmark with 4KB random read running on CPU 1 increases IOPS by 58%:

  Name              IOPS   Error
  Before        26720.59 =C2=B1 0.28%
  After         42373.79 =C2=B1 0.54%

Now most of this improvement is not due to NUMA but just because the requests
complete on the same CPU where they were submitted. However, if the IRQ is on
CPU 0 and fio also runs on CPU 0 only 39600 IOPS is achieved, not the full
42373 IOPS that we get when NUMA affinity is honored. So it is worth taking
NUMA into account to achieve maximum performance.

The following patches are a hack that uses the device's NUMA node when
assigning managed IRQs. They are not mergeable but I hope they will help start
the discussion. One bug is that they affect all managed IRQs, even for devices
with many IRQs where spreading across all CPUs is a good policy.

Please let me know what you think:

1. Is there a reason why managed IRQs should *not* take NUMA into account that
   I've missed?

2. Is there a better place to implement this logic? For example,
   pci_alloc_irq_vectors_affinity() where the cpumasks are calculated.

Any suggestions on how to proceed would be appreciated. Thanks!

Stefan Hajnoczi (2):
  genirq: honor device NUMA node when allocating descs
  genirq/matrix: take NUMA into account for managed IRQs

 include/linux/irq.h           |  2 +-
 arch/x86/kernel/apic/vector.c |  3 ++-
 kernel/irq/irqdesc.c          |  3 ++-
 kernel/irq/matrix.c           | 16 ++++++++++++----
 4 files changed, 17 insertions(+), 7 deletions(-)

--=20
2.26.2