lllqos: Linux Low-Latency QoS for Python

The Linux kernel provides the PM Quality Of Service Interface, which allows userspace processes to indicate acceptable wake-up latency. Processes seeking low wake-up latency (e.g., cyclictest) use this interface to tell the kernel the acceptable target latency (in microseconds).

The kernel documentation describes the interface as follows:

To register the default PM QoS target for the CPU latency QoS, the process must open /dev/cpu_dma_latency.

As long as the device node is held open that process has a registered request on the parameter.

To change the requested target value, the process needs to write an s32 value to the open device node.

This package provides a simple context-manager API low_wakeup_latency(target_value) that automates this. By default, the target value is 0.

On platforms other than Linux (or if the /dev filesystem is not mounted in the usual place), this context manager has no effect.

Example

from lllqos import low_wakeup_latency

# Standard use:
with low_wakeup_latency():
    print('Hi, but quickly!')

# Check whether lease was acquired:
with low_wakeup_latency() as lease:
    print(f'Lease acquired: {lease.acquired()}')
    
# Setting some nonzero goal:
with low_wakeup_latency(1000):
    print('Up to 1ms latency is acceptable here.')

A typical use-case is to combine this package with the abs-sleep package to achieve low-latency periodic activations.

from lllqos import low_wakeup_latency
from abs_sleep import periodic
import time

# tell Linux to minimize wake-up latency
with low_wakeup_latency():
  	# periodic wake-ups with a period of 500ms
    for t in periodic(0.5, align_secs=1):
      	now = time.monotonic()
      	print(f"periodic activation @ {t}s with {(now - t) * 1000000:.2f}µs latency")

NB: To achieve consistently low latency, you must run the Python interpreter with real-time priority (e.g., with theSCHED_FIFO or SCHED_RR scheduling policies).

See Also

• Linux kernel documentation
• cyclictest implementation