Really interesting benchmark, thank you.

> On 8 Nov 2023, at 23:04, Martin Desruisseaux 
> <[email protected]> wrote:
> 
> Hello all
> 
> Apache SIS 1.4 contains a Cloud Optimized GeoTIFF reader, and a GeoTIFF 
> writer is currently available in 1.5-SNAPSHOT (not yet fully completed). 
> Those reader/writer are in pure Java (ignoring native code provided by the 
> JDK itself). There is a widespread belief that a Java code would surely be 
> slower than C/C++ code. To verify, we conducted a superficial benchmark. It 
> is very superficial in that we tested only one image with one compression 
> method, no sub-region, no subsampling, no parallelization, no reprojection, 
> etc. But the results nevertheless question the above-cited belief. In that 
> benchmark, Apache SIS and GDAL performances were equivalent. The numbers 
> actually report Apache SIS as very slightly faster than GDAL, but we cannot 
> conclude much because of all above-cited limitations, because there is a 
> possibility of biais in time measurements (we tried to compensate it), 
> because the differences are close to margin errors, and because Apache SIS 
> spent 95% of its time in the native code of the `java.util.zip` package. 
> Assuming that GDAL also uses the `zlib` library (we did not tested with 
> `libdeflate`), 95% of this benchmark is actually measuring the same code. 
> However for the remaining 5%, it suggests that the interfacing between Java 
> and the native code in `java.nio` and `java.util.zip` packages can be as 
> efficient as the interfacing between a C/C++ application and native libraries.
> 
> More details are there: 
> https://www.geomatys.com/2023/10/05/geotiff-reader-writer-performance-comparison/
> 
>     Martin
> 
> 

Attachment: signature.asc
Description: Message signed with OpenPGP

Reply via email to