OTB prefetch

This remote module of the Orfeo ToolBox provides a new application to prefetch a pipeline's input in an asynchronous way.

Under the hood, the otb::PrefetchCacheAsyncFilter does the work. It tries to guess the next requested region, and request it to the upstream pipeline, while the downstream pipeline runs on its own.

Example

In a deep learning application, at inference time, avoiding the extra cost of the GPU idle while GDAL is just grabbing the next chunk of image (example with pyotb):

import pyotb

img_href = "https://.../input.tif"
cal = pyotb.OpticalCalibration(img_href)
prefetch = pyotb.Prefetch(cal)
infer = pyotb.TensorflowModelServe(prefetch, ...)
infer.write("output.tif")

After the execution, the filter display some performance metrics:

PrefetchCacheAsyncFilter (0x563812708470): 7.0272e+06 missing guessed pixels (4.68247 %)
PrefetchCacheAsyncFilter (0x563812708470): 1.43047e+08 good guessed pixels (95.3175 %)
PrefetchCacheAsyncFilter (0x563812708470): 351360 extra guessed pixels (0.234123 %)
PrefetchCacheAsyncFilter (0x563812708470): Total wait: 0.000963238s

We can see that less that 1ms has bee spent between the GPU calls, saving us some money! (for information the input is a Sentinel-2 file in COG format from Microsoft Planetary Computer, processed with a streaming:type=tiled strategy on the OTB writer). If you want to know more about OTB/GDAL writing strategies, this is a good read.

Build from source

You have to build the remote module from sources. Give a try with one otbtf development docker image:

cd /src/otb/otb/Modules/Remote
git clone https://github.com/remicres/prefetch
cd /src/otb/build/OTB/build
cmake /src/otb/otb -DModule_OTBPrefetch=ON
make install

You can then check that the OTB application has been installed:

otbcli_Prefetch --help

License

Apache 2.0

Contact

Remi Cresson at INRAE

Acknowledgements

• Bradley Lowekamp
• OTB Dev team