What is NVMe over TCP

NVMe over TCP is like iSCSI, a standard for network-based storage. It provides block-level access to NVMe drives over TCP/IP, having much higher performance than iSCSI.

iSCSI was originally designed for SCSI drives, i.e. SSD (Slow Spinning Disks). It’s an old protocol that does not suit the needs of today’s SSDs.

The kernel Linux has native support for NVMe over TCP.

My setup

Motherboard: Supermicro H12SSL-NT
RAM: plenty
Storage: 4 x 7.68TB Kioxia CD6 U.3 NVMe SSD, ZFS RAIDz1
Network: Onboard Broadcom 10GbE
OS: Debian 12 (bookworm)

Motherboard: a good one
CPU: AMD Ryzen 9 5950X
RAM: plenty
Network: TP-link AQC107 10GbE
OS: Arch Linux, kernel 6.3.y

They are connected to the switch ports of my home router (OpenWRT x86_64 on Ryzen 3 3100 with Intel X710 10GbE, promiscuous mode on).

MTU is set to 9000.

Configure target

  • Install nvme-cli package from the distro’s package manager, if not already installed.

  • Load nvmet-tcp kernel module if it is not loaded at boot:

modprobe nvmet-tcp

Alternatively, let the kernel load nvmet-tcp kernel module at boot:

echo "nvmet_tcp" > /etc/modules-load.d/nvmet_tcp.conf
  • Create and configure an NVMe target subsystem, let’s call it “mysub”:
mkdir /sys/kernel/config/nvmet/subsystems/mysub
cd /sys/kernel/config/nvmet/subsystems/mysub
echo 1 > attr_allow_any_host

If the kernel module nvmet-tcp is loaded, directory /sys/kernel/config/nvmet should exist.

If running as sudo, echo 1 > attr_allow_any_host might fail. In this case, su root or try echo 1 | sudo tee -a attr_allow_any_host > /dev/null instead.

  • Before creating a namespace for the target, use lsblk or nvme list to find out the name of the NVMe device to be attached to the target (i.e. /dev/nvme0n1).

  • Next, create and configure a namespace 1:

mkdir /sys/kernel/config/nvmet/subsystems/mysub/namespace/1
cd /sys/kernel/config/nvmet/subsystems/mysub/namespace/1
echo -n /dev/nvme0n1 > device_path
echo 1 > enable
  • Configure a port 1 for an initiator to connect:
mkdir /sys/kernel/config/nvmet/ports/1
cd /sys/kernel/config/nvmet/ports/1
echo "ipv4" > addr_adrfam
echo "tcp" > addr_trtype
echo 4420 > addr_trsvcid
echo > addr_traddr

Here, our target has a static IPv4 address 4420 is the common port number used for NVMe-oF connection.

  • Create a symbolic link from the mysub subsystem to the port 1
ln -s /sys/kernel/config/nvmet/subsystems/mysub/ /sys/kernel/config/nvmet/ports/1/subsystems/mysub

Using dmesg | grep nvme_tcp, we can see in the kernel log that the port is enabled.

Configure initiator

  • Install nvme-cli package from the distro’s package manager, if not already installed.

  • Load nvme-tcp kernel module (NOT nvmet-tcp) if it is not loaded at boot:

modprobe nvme-tcp

Alternatively, let the kernel load nvme-tcp kernel module at boot:

echo "nvme_tcp" > /etc/modules-load.d/nvme_tcp.conf
  • Attempt to discover a remote target. This step can be skipped.
nvme discover -t tcp -a -s 4420
  • Connect the NVMe device:
nvme connect -t tcp -a -s 4420 -n mysub

The subsystem NQN argument -n mysub can be replaced by a host NQN, which is stored in the target’s /etc/nvme/hostnqn. The command will look like this:

nvme connect -t tcp -a -s 4420 --hostnqn=nqn.2014-08.org.nvmexpress:uuid:1b4e28ba-2fa1-11d2-883f-0016d3ccabcd
  • Check nvme list or lsblk for the attached device. To disconnect, first unmount the mounted NVMe device, then run nvme disconnect:
nvme disconnect /dev/nvme2n1 -n mysub


Definitely faster than NFS, Samba, or iSCSI.

After formatting the zvol to ext4, I was able to use the package fio-3.34 to test the I/O performance. Here is the test profile I used:




I got about 650MB/s sequential read/write and about 90K IOPS random read/write, slightly better than a typical SATA SSD. The numbers were not great, probably because of the suboptimal zpool setup. Mounting the zvol on the target directly, I only got about 1.5GB/s sequential read. Also, the ethernet adapter at the initiator’s side, AQC107, is known to be not very stable and sometimes overheat (and no support for RDMA). It would be better if I had a 25Gb or 40Gb network adapter.