-
Notifications
You must be signed in to change notification settings - Fork 14
Basecalling settings
Miles edited this page Dec 7, 2020
·
3 revisions
| ❗ This section is under construction |
|---|
General formula:
runners * chunks_per_runner * chunk_size < 100'000 * [max GPU memory in GB] * 2
So the current settings I am running on the Xavier NX are:
- --gpu_runners_per_device = 4 (1 default MinIT config)
- --chunks_per_runner = 256 (48 default MinIT config)
- --chunk_size = 1000 (1000 default MinIT config)
4 * 256 * 1000 ~= 100'000 * 8 * 2 which is 1'024'000 < 1'600'000
this is leaving some overhead for the likes of the OS and MinKnow.
Running with the default settings (as above) took ~5.3 mins on a test data set. Running with the selected settings reduced this down to ~2.7 mins.
Increasing the chunks_per_runner led to the largest gains in speed.
The current 'optimal' settings I am running on the Xavier AGX are:
- --num_callers (num basecallers) = 2
- --gpu_runners_per_device = 3 (1 default MinIT config)
- --chunks_per_runner = 1024 (48 default MinIT config)
- --chunk_size = 1000 (1000 default MinIT config)
3 * 1024 * 1000 ~= 100'000 * 16 * 2 which is 3'072'000 < 3'200'000
This is giving me a base calling rate of ~9.43x106 samples/s.