[foreign] Poor performance?

[Maurizio Cimadamore]

On 17/05/2019 06:54, Samuel Audet wrote:
> Hi,
>
> This person seems to be unable to obtain an acceptable level of performance
> with Panama:
> https://github.com/zakgof/java-native-benchmark
>
> I thought it would be good to know why so this can be fixed!
Hi Samuel,

thanks you for bringing this up, I saw this benchmark few days ago and I 
took a look at it. That benchmark is unfortunately hitting on a couple 
of (transitory!) pain points: (1) it is running on Windows, which lacks 
the optimizations available for MacOS and Linux (directInvoker). When 
the linkToNative effort will be completed, this discrepancy between 
platforms will go away. The second problem (2) is that the call is 
passing a big struct (e.g. bigger than 64 bits). Even on Linux and Mac, 
such a call would be unable to take advantage of the optimized invoker 
and would fall back to the so called 'universal invoker' which is slow.

The plan is to start using linkToNative as our official cross-platform 
fast-path, and then progressively enhance linkToNative so that it can 
handle all the cases that are currently left to universalInvoker - of 
these there are two that are important: big structs passed by pointer, 
and return value in memory. Once this is done (and this is mostly a 
matter of figuring out a 'private protocol', or set of carriers that are 
known to both the Panama binder and the VM), then I would expect that 
performance cliffs like the one you show will go away.

At present state, on Linux and Mac, I would expect a call involving 
small structs and/or primitive to be on par with JNI and be 
significantly faster than JNI/JFFI. Luckily, when writing real world 
code, these things usually cancel out - e.g. it would be unusual for a 
real world application to only have 'slow' calls with big structs, but 
the situation is of course possible in a synthetic benchmark and we 
should make sure to polish these edges.

Thanks
Maurizio

>
> Samuel