Amazon EBS and NVMe on Linux instances

EBS volumes are exposed as NVMe block devices on instances built on the Nitro System. The device names are /dev/nvme0n1, /dev/nvme1n1, and so on. The device names that you specify in a block device mapping are renamed using NVMe device names (/dev/nvme[0-26]n1). The block device driver can assign NVMe device names in a different order than you specified for the volumes in the block device mapping.

The EBS performance guarantees stated in Amazon EBS Product Details are valid regardless of the block-device interface.

Install or upgrade the NVMe driver

To access NVMe volumes, the NVMe drivers must be installed. Instances can support NVMe EBS volumes, NVMe instance store volumes, both types of NVMe volumes, or no NVMe volumes. For more information, see Summary of networking and storage features.

The following AMIs include the required NVMe drivers:

For more information about NVMe drivers on Windows instances, see Amazon EBS and NVMe on Windows Instances in the Amazon EC2 User Guide for Windows Instances.

To confirm that your instance has the NVMe driver

You can confirm that your instance has the NVMe driver and check the driver version using the following command. If the instance has the NVMe driver, the command returns information about the driver.

$ modinfo nvme

Identifying the EBS device

EBS uses single-root I/O virtualization (SR-IOV) to provide volume attachments on Nitro-based instances using the NVMe specification. These devices rely on standard NVMe drivers on the operating system. These drivers typically discover attached devices by scanning the PCI bus during instance boot, and create device nodes based on the order in which the devices respond, not on how the devices are specified in the block device mapping. In Linux, NVMe device names follow the pattern /dev/nvme<x>n<y>, where <x> is the enumeration order, and, for EBS, <y> is 1. Occasionally, devices can respond to discovery in a different order in subsequent instance starts, which causes the device name to change.

We recommend that you use stable identifiers for your EBS volumes within your instance, such as one of the following:

Amazon Linux AMIs

With Amazon Linux AMI 2017.09.01 or later (including Amazon Linux 2), you can run the ebsnvme-id command as follows to map the NVMe device name to a volume ID and device name:

[ec2-user ~]$ sudo /sbin/ebsnvme-id /dev/nvme1n1
Volume ID: vol-01324f611e2463981
/dev/sdf

Amazon Linux also creates a symbolic link from the device name in the block device mapping (for example, /dev/sdf), to the NVMe device name.

FreeBSD AMIs

Starting with FreeBSD 12.2-RELEASE, you can run the ebsnvme-id command as shown above. Pass either the name of the NVMe device (for example, nvme0) or the disk device (for example, nvd0 or nda0). FreeBSD also creates symbolic links to the disk devices (for example, /dev/aws/disk/ebs/volume_id).

Other Linux AMIs

With a kernel version of 4.2 or later, you can run the nvme id-ctrl command as follows to map an NVMe device to a volume ID. First, install the NVMe command line package, nvme-cli, using the package management tools for your Linux distribution. For download and installation instructions for other distributions, refer to the documentation specific to your distribution.

The following example gets the volume ID and device name. The device name is available through the NVMe controller vendor-specific extension (bytes 384:4095 of the controller identification):

[ec2-user ~]$ sudo nvme id-ctrl -v /dev/nvme1n1
NVME Identify Controller:
vid     : 0x1d0f
ssvid   : 0x1d0f
sn      : vol01234567890abcdef
mn      : Amazon Elastic Block Store
...
0000: 2f 64 65 76 2f 73 64 6a 20 20 20 20 20 20 20 20 "/dev/sdf..."

The lsblk command lists available devices and their mount points (if applicable). This helps you determine the correct device name to use. In this example, /dev/nvme0n1p1 is mounted as the root device and /dev/nvme1n1 is attached but not mounted.

[ec2-user ~]$ lsblk
NAME          MAJ:MIN RM SIZE RO TYPE MOUNTPOINT
nvme1n1       259:3   0  100G  0 disk
nvme0n1       259:0   0    8G  0 disk
  nvme0n1p1   259:1   0    8G  0 part /
  nvme0n1p128 259:2   0    1M  0 part

Working with NVMe EBS volumes

To format and mount an NVMe EBS volume, see Making an Amazon EBS volume available for use on Linux.

If you are using Linux kernel 4.2 or later, any change you make to the volume size of an NVMe EBS volume is automatically reflected in the instance. For older Linux kernels, you might need to detach and attach the EBS volume or reboot the instance for the size change to be reflected. With Linux kernel 3.19 or later, you can use the hdparm command as follows to force a rescan of the NVMe device:

[ec2-user ~]$ sudo hdparm -z /dev/nvme1n1

When you detach an NVMe EBS volume, the instance does not have an opportunity to flush the file system caches or metadata before detaching the volume. Therefore, before you detach an NVMe EBS volume, you should first sync and unmount it. If the volume fails to detach, you can attempt a force-detach command as described in Detaching an Amazon EBS volume from a Linux instance.

I/O operation timeout

EBS volumes attached to Nitro-based instances use the default NVMe driver provided by the operating system. Most operating systems specify a timeout for I/O operations submitted to NVMe devices. The default timeout is 30 seconds and can be changed using the nvme_core.io_timeout boot parameter. For most Linux kernels earlier than version 4.6, this parameter is nvme.io_timeout.

If I/O latency exceeds the value of this timeout parameter, the Linux NVMe driver fails the I/O and returns an error to the filesystem or application. Depending on the I/O operation, your filesystem or application can retry the error. In some cases, your filesystem might be remounted as read-only.

For an experience similar to EBS volumes attached to Xen instances, we recommend setting nvme_core.io_timeout to the highest value possible. For current kernels, the maximum is 4294967295, while for earlier kernels the maximum is 255. Depending on the version of Linux, the timeout might already be set to the supported maximum value. For example, the timeout is set to 4294967295 by default for Amazon Linux AMI 2017.09.01 and later.

You can verify the maximum value for your Linux distribution by writing a value higher than the suggested maximum to /sys/module/nvme_core/parameters/io_timeout and checking for the Numerical result out of range error when attempting to save the file.