VRT Performance Benchmarking

2022-11-10

Overview

Test reads elevation value from ArcticDem. POI is lon: -74.60 lat: 82.86
Method used vrt file created from mosaic rasters, version 3.0, 2m, hosted on AWS.
The mosaic.vrt file is stored locally on aws dev server at /data/ArcticDem/mosaic.vrt

The actual raster containing the elevation for POI is:
/vsis3/pgc-opendata-dems/arcticdem/mosaics/v3.0/2m/34_37/34_37_1_1_2m_v3.0_reg_dem.tif

The elevation can be read from the aws file directly:
gdallocationinfo -wgs84 /vsis3/pgc-opendata-dems/arcticdem/mosaics/v3.0/2m/34_37/34_37_1_1_2m_v3.0_reg_dem.tif -valonly -74.6 82.86

The elevation can be read from the mosaic.vrt on our dev server:
gdallocationinfo -wgs84 /data/ArcticDem/mosaic.vrt -valonly -74.6 82.86

The test reads the same POI one million times. Gdal library should access the proper raster tile from S3 bucket and use it locally for all remaining elevation reads.

In the first implementation the mosaic.vrt is opened, the vrtdataset and vrtband are used to do a direct read via:
vrtband->RasterIO(GF_Read, col, row, 1, 1, &elevation, 1, 1, GDT_Float32, 0, 0, 0);
col and row are calculated for the mosaic.vrt

This approach works correctly but performance is very poor. It took almost 170 seconds to do one million reads. Each 100k reads takes almost 17 seconds.

Points read: 100000 16988.167999778
Points read: 200000 16991.582000162
Points read: 300000 16994.273000164
Points read: 400000 16995.032000123
Points read: 500000 16998.428999912
Points read: 600000 16988.674999913
Points read: 700000 16993.327999953
Points read: 800000 17000.902000116
Points read: 900000 16991.788999876
Points read: 1000000 16998.240999877

1000000 points read time: 169940.77899982 msecs

In the second approach the vrtdataset and vrtband are used ONLY to find the name of the raster containing POI.
The code is very similar to how gdallocationinfo tool is implemented using xml parsing in GDAL.
The raster tif file is than opened and this raster’s dataset and band are used to do a read using col, row calculated for that raster. This method is much more efficient. It only took 665 msecs to do one million reads.

Points read: 100000 66.570000024512
Points read: 200000 66.518999869004
Points read: 300000 66.537000006065
Points read: 400000 66.584999905899
Points read: 500000 66.590999951586
Points read: 600000 66.584999905899
Points read: 700000 66.558999940753
Points read: 800000 66.634999820963
Points read: 900000 66.579000093043
Points read: 1000000 66.580000100657

1000000 points read time: 665.83800013177 msecs

The test was also repeated with POI beigng different each time read is executed but always within the same raster file. This caused more S3 transactions for GDAL to read many tiles. The performance degraded where the second ‘fast’ method took almost 2 seconds to read a million POIs which the first method took almost the same amount of time.