Multipath configuration for FC-NVMe hosts

Follow FC-NVMe multipath configuration recommendations for successful attachment of Linux® hosts to the system.

Hosts can be configured to work with the traditional Device Mapper or with the Native NVMe Multipath. For SCSI devices, the host continues to work with Device Mapper in any case. Native NVMe Multipath is supported only on SLES12SP4 and on SLES15. Native NVME Multipath is not supported on Red Hat Enterprise Linux. Native NVMe Multipath supports NVMe ANA (the NVMe equivalent to SCSI ALUA), whereas Device Mapper currently does not support ANA. In Native NVMe Multipath, every namespace is shown as one instance.

Turning Native NVMe Multipath on or off

On SLES15, Native NVMe Multipath is turned on by default. On SLES12SP4, Native NVMe Multipath is not turned on by default.

To check whether Native NVMe Multipath is on, enter:
# systool -m nvme_core -A multipath
Module = "nvme_core"

    multipath           = "Y"
If the multipath = "N", and you want to turn it on, enter:

echo "options nvme_core multipath=Y" > /etc/modprobe.d/50-nvme_core.conf
dracut -f
reboot
If you choose to work with Device Mapper, enter:
echo "options nvme_core multipath=N" > /etc/modprobe.d/50-nvme_core.conf
dracut -f
reboot

Device Mapper or Native NVMe Multipath configuration

Two framework options are available to configure the multipath configuration file: Device Mapper or Native NVMe Multipath.

To configure Device Mapper, complete the following steps:
  1. Edit the /etc/multipath.conf file to include the following code:
    devices {
        device {
            vendor "NVME"
            product "IBM     2145"
            path_grouping_policy "multibus"
            path_selector "round-robin 0"
            prio "const"
            path_checker "none"
            failback "immediate"
            no_path_retry "queue"
            rr_weight uniform
            rr_min_io_rq "1"
            fast_io_fail_tmo 15
            dev_loss_tmo 600
        }
    }
    defaults {
        user_friendly_names yes
        path_grouping_policy    group_by_prio
    }
  2. Run the following commands to validate that the multipath daemon is running:
    systemctl enable multipathd.service
    systemctl start multipathd.service
    
    # ps -ef | grep -v grep | grep multipath
    root      1616     1  0 Nov21 ?        00:01:14 /sbin/multipathd -d -s
    
    After you enable the multipath service for SLES12SP4, rebuild initrd with multipath support:
    dracut --force --add multipath
  3. Run the following commands to apply configurations:
    multipath -F
    multipath
    multipath -ll
If you use Native NVMe Multipath, add the following code to the multipath configuration file:
blacklist {
    device {
        vendor  "NVME"
        product "IBM\s+2145"
    }
}

Device Mapper performance

Testing shows that performance is superior when you work with multiqueue I/O scheduling as described in the following steps.
Note: Before you apply these steps, contact a SUSE representative to verify that this option is supported.
For performance improvements, apply multiqueue I/O scheduling with blk-mq (in SLES): 
  1. Edit the /etc/default/grub file and the following text:
    GRUB_CMDLINE_LINUX_DEFAULT="BOOTPTimeout=20 BootpWait=20 
    biosdevname=0 powersaved=off resume=/dev/system/swap splash=silent 
    quiet showopts crashkernel=175M,high dm_mod.use_blk_mq=y 
    scsi_mod.use_blk_mq=1 transparent_hugepage=never"
  2. Apply the new configuration:
    swfc178:~ # grub2-mkconfig -o /boot/grub2/grub.cfg
    Generating grub configuration file ...
    Found theme: /boot/grub2/themes/SLE/theme.txt
    Found linux image: /boot/vmlinuz-4.12.14-25.19-default
    Found initrd image: /boot/initrd-4.12.14-25.19-default
    Found linux image: /boot/vmlinuz-4.12.14-23-default
    Found initrd image: /boot/initrd-4.12.14-23-default
    done
  3. Reboot.
  4. Validate that the multiqueue feature is enabled in multipath:
    mpatho (eui.880000000000000b0050760071c60044) dm-3 NVME,IBM     2145
    size=1.5G features='3 queue_if_no_path queue_mode mq' hwhandler='0' wp=rw