24.1 beta 23.3 23.2 23.1

rpk redpanda tune list

This command is not supported on macOS, Windows.

List available Redpanda tuners and check if they are enabled and supported by your system.

To enable a tuner it must be set in the redpanda.yaml configuration file under rpk section, for example:

rpk:
tune_cpu: true
tune_swappiness: true

You may use rpk redpanda config set to enable or disable a tuner.

Usage

rpk redpanda tune list [flags]

Flags

Value Type Description

-r, --dirs

strings

List of data directories or places to store data (e.g. /var/vectorized/redpanda/); usually your XFS file system on an NVMe SSD device.

-d, --disks

strings

Lists of devices to tune f.e. 'sda1'.

-h, --help

-

Help for list.

-m, --mode

string

Operation Mode: one of: [sq, sq_split, mq].

-n, --nic

strings

Network Interface Controllers to tune.

--reboot-allowed

-

Allow tuners to tune boot parameters and request system reboot.

--config

string

Redpanda or rpk config file; default search paths are /var/lib/redpanda/.config/rpk/rpk.yaml, $PWD/redpanda.yaml, and /etc/redpanda/redpanda.yaml.

-X, --config-opt

stringArray

Override rpk configuration settings. See rpk -X or execute rpk -X help for inline detail or rpk -X list for terser detail.

--profile

string

Profile to use. See rpk profile for more details.

-v, --verbose

-

Enable verbose logging.

Example output

The output of the command lists each available tuner and whether it’s enabled or supported (with a reason if it’s unsupported).

TUNER                  ENABLED  SUPPORTED  UNSUPPORTED-REASON
aio_events             true     true
ballast_file           true     true
clocksource            true     false      Clocksource setting not available for this architecture
coredump               false    true
cpu                    true     true
disk_irq               true     true
disk_nomerges          true     true
disk_scheduler         true     false      open : no such file or directory
disk_write_cache       true     false      Disk write cache tuner is only supported in GCP
fstrim                 false    false      dial unix /var/run/dbus/system_bus_socket: connect: no such file or directory
net                    true     true
swappiness             true     true
transparent_hugepages  false    true

Tuners

This section describes each tuner and their configuration settings in redpanda.yaml. It provides overlapping and additional content to the descriptions returned by rpk redpanda tune help.

aio_events

The aio_events tuner maximizes throughput by increasing the number of simultaneous I/O requests. It sets the maximum number of outstanding asynchronous I/O operations to be above a calculated threshold.

Configuration (yaml):

  • rpk.tune_aio_events: flag to enable the tuner (Default for production mode: true)

ballast_file

The ballast_file tuner improves debuggability and recoverability of a system with a full disk by creating a ballast file: if the disk becomes full, the ballast file can be deleted, thereby unblocking processes that are failing due to write failures and providing time for operators to stop any problematic processes causing the disk to fill, and to delete topics or records and change retention settings.

Configuration (yaml):

  • rpk.tune_ballast_file: flag to enable the tuner (Default for production mode: true)

  • rpk.ballast_file_size: size of the ballast file in bytes (Default for production mode: 1 GiB)

  • rpk.ballast_file_path: path to the ballast file (Default for production mode: /var/lib/redpanda/data/ballast)

clocksource

The clocksource tuner improves the performance of getting system time by changing the clock source. It sets the clock source to the Time Stamp Counter (TSC) so time can be read in userspace via the Virtual Dynamic Shared Object (vDSO). This is more efficient than the usual virtual machine clock source, xen, that involves the overhead of a system call.

Configuration (yaml):

  • rpk.tune_clocksource: flag to enable the tuner (Default for production mode: true)

coredump

The coredump tuner enables coredumps to be saved for production mode. By default, coredumps are not enabled for production. Enabling this tuner sets the location to save coredumps.

Configuration (yaml):

  • rpk.tune_coredump: flag to enable the tuner (Default for production mode: false)

  • rpk.coredump_dir: path to the saved coredump (Default for production mode: /var/lib/redpanda/coredump)

cpu

The cpu tuner maximizes CPU performance for I/O intensive workloads by optimizing various CPU settings. It performs the following operations:

  • Sets the ACPI-cpufreq governor to performance

If system reboot is allowed (rpk redpanda tune --reboot-allowed option is set), the system must be rebooted after the first pass of the tuner, and it performs additional operations:

  • Disables Intel P-states

  • Disables Intel C-states

  • Disables turbo boost

After tuning, the system CPUs operate at the maximum non-turbo frequency.

Configuration (yaml):

  • rpk.tune_cpu: flag to enable the tuner (Default for production mode: true)

disk_irq

The disk_irq tuner optimizes the handling of interrupt requests (IRQs) for disks binding all disk IRQs to a requested set of CPUs. It tries to distribute IRQs according to the following guidelines:

  • Distribute NVMe disks IRQs equally among all available CPUs.

  • Distribute non-NVMe disks IRQs equally among designated CPUs or among all available CPUs in the mq mode.

IRQs are distributed according to the operation mode set by rpk redpanda tune --mode <operation-mode>. The available operation modes:

  • sq: set all IRQs of a given device to CPU0

  • sq_split: divide all IRQs of a given device between CPU0 and its HT siblings

  • mq: distribute device IRQs among all available CPUs instead of binding them all to CPU0

If no --mode is specified, a default mode is determined:

  • If there are only NVMe disks, the mq mode is set as the default.

  • For non-NVMe disks:

    • If the number of HT siblings is less than or equal to four, the mq mode is set as the default.

    • Otherwise, if the number of cores is less than or equal to four, the sq mode is set as the default.

    • For all other conditions, the sq_split mode is set as the default.

Configuration (yaml):

  • rpk.tune_disk_irq: flag to enable the tuner (Default for production mode: true)

  • rpk redpanda tune --mode <operation-mode> sets the IRQ distribution mode

net

The net tuner optimizes the handling of interrupt requests (IRQs) for network interfaces (NICs) by binding all NIC IRQs to a requested set of CPUs.

Its IRQ distribution operation modes are the same as described for the disk_irq tuner with NICs as the devices.

Configuration (yaml):

  • rpk.tune_network: flag to enable the tuner (Default for production mode: true)

  • rpk redpanda tune --mode <operation-mode> sets the IRQ distribution mode

disk_nomerges

The disk_nomerges tuner reduces CPU overhead by disabling the merging of adjacent I/O requests.

Configuration (yaml):

  • rpk.tune_disk_nomerges: flag to enable the tuner (Default for production mode: true)

disk_scheduler

The disk_scheduler tuner optimizes disk scheduler performance for the type of device (NVME, non-NVME). It provides a selectable set of schedulers:

  • none: minimizes latency of modern NVMe devices by bypassing the operating system’s I/O scheduler

  • noop: preferred for non-NVME devices (and used when none is unavailable), this scheduler uses a simple FIFO queue where all I/O operations are first stored and then handled by the driver.

Configuration (yaml):

  • rpk.tune_disk_scheduler: flag to enable the tuner (Default for production mode: true)

disk_write_cache

The disk_write_cache tuner optimizes performance in Google Cloud Platform (GCP) by enabling write-through caching for its NVMe Local SSD drives.

Configuration (yaml):

  • rpk.tune_disk_write_cache: flag to enable the tuner (Default for production mode: true)

fstrim

The fstrim tuner improves SSD performance by starting a background systemd service to periodically wipe memory blocks that are not used by the file system. This is desirable for SSDs because they require wiping the space where new data will be written, so not wiping during non-write cycles will eventually cause performance degradations, when the lack of free space results in writes triggering synchronous erasures.

If it’s available, the fstrim systemd service will be run. If it’s unavailable but systemd is available, an equivalent service will be installed and run. Otherwise, no service will be run.

Configuration (yaml):

  • rpk.tune_fstrim: flag to enable the tuner (Default for production mode: true)

swappiness

The swappiness tuner tunes the kernel to keep process data in-memory for as long as possible instead of swapping it out to disk.

Configuration (yaml):

  • rpk.tune_swappiness: flag to enable the tuner (Default for production mode: true)

transparent_hugepages

Redpanda makes excellent use of operating system-level Transparent HugePages (THP) by pre-allocating memory at startup. For CPUs that support THPs, this results in improved memory page caching and reduced cache misses from translation lookaside buffer (TLB) lookups.

Most operating systems configure THP use to madvise by default, meaning a process must explicitly request them. Redpanda is designed to make the required madvise call, meaning THPs are in use under typical operation scenarios. Other possible values for THP use are disabled or always.

The transparent_hugepages tuner overrides the operating system setting, configuring THP use to always. This tuner only has a practical effect if the operating system is configured to disabled for THP use. For this reason, this tuner is disabled by default in production mode. It primarily exists to allow operators to ensure THPs are in use for their installations if there is any possibility the operating system is configured to disabled.

Configuration (yaml):

  • rpk.tune_transparent_hugepages: flag to enable the tuner (Default for production mode: false)