2024-01-10 Lumiera Developers Meeting

Once the scheduler had been implemented it was necessary to subject it to various load tests to establish its stability, and stress tests were necessary to provide information on performance limitations so that we would be able to establish whether these limitations met the demands required by Lumiera. Results were encouraging.

Ichthyo reported on stress tests on the new scheduler. After resolving a series of initial bugs, processing runs stable now, even when entering over 500 test jobs simultaneously to the queue. Stress tests revealed typical limitations of the scheduler, which exhibited behaviour that is well within the requirements of Lumiera. Load tests showed that the scheduler remained stable when subjected to short heavy load bursts within expected bounds.

Indeed performance testing is always a delicate and complex area to investigate, as it is necessary to establish reasonable performance boundaries of the system: we can always overload any scheduler with fanatically huge numbers of jobs; while the scheduler is still able to process all enqueued jobs eventually, congestion builds up and jobs start missing their deadlines. Of interest, however, is to establish some reasonable information on the processing rate of the scheduler which Ichthyo was able to estimate statistically by carefully observing the scheduler under a collection of pre-calibrated loads.

Some goals of the tests:

Tests on a wider array of hardware constellations are not expected to reveal much relevant information in this stage of the project and will thus be conducted later. This kind of analytical, isolated component tests will not ultimately resolve the issue of whether the scheduler can perform as desired, but subjecting the scheduler to carefully selected tests can help to reduce the probability of missing serious performance limitations inherent to the chosen design.

Benn has written a simple GStreamer demo so that a video in raw format can be displayed on the GUI for the current vertical slice. The next step in this investigation is to identify a low-level entrance point to the functionality provided by the GStreamer library, since for use in this vertical slice the render engine produces a buffer filled with raw video data for each frame, which will then be displayed in a video widget as an ongoing process.

This is a general question requiring further investigation: how to make use of established functionality from a library to simplify our implementation task, while retaining the full control and coordination of the rendering and playback process and resource management within the Lumiera render engine itself.

The next Lumiera meeting is scheduled for Wednesday, Feb 14 at 20:00 UTC. Should this be inconvenient for you, please speak up on, for example, lumiera@lists.lumiera.org and let us know of when you would prefer to attend.