On the performance and expressiveness of the PPL java bindings
Hi all, I am using the PPL library from Java (hmm.. Scala, actually) to make experiments on abstract interpretation based static analysis. I would like to share some considerations about a couple of aspects of the Java bindings which I believe could be improved.
First aspect is performance. Obviously, using native libraries in the JVM always incurs a certain amount of performance penalty. I tried to develop a native implementation of Double_Box, and it is much faster than using the Double_Box fron PPL (the native implementation is not safe since I cannot control rounding mode within Java, but I do not think this impacts performance a lot). Obviously for more complex domains the overhead of calling native methods is proportionally lower.
My question is: is it possible to improve performance of the Java bindings? I am not an expert in JNI, but from a fast scan of the code, it seems the Java code passes to the native methods an object. The native code calls some accessor function to get the ptr field (which is the pointer to the native PPL object). It should be possible to rewrite the native functions so that they receive directly the pointer, and therefore do not need to call the accessor methods. From what I have heard, calling accessor methods is slow. Has this approach been pursued and abandoned or not? Do you think it could produce a noticeable improvement in performance?
The other point regards expressiveness. Java classes in PPL form a flat hierarchy. This makes difficult to write generic code, parametric w.r.t. the abstract domain. In C++ there is a flat hierarchy, too, but there templates may be used to write generic code effectively. In Java, the only choices to write generic code are either to use reflection (which is quite costly) or to write thin wrappers over the PPL classes, and put wrappers in a hierarchy, (which is tedious). Who works in Scala may use macros, but in any case, writing generic code is cumbersome, and the model of a flat hierarchy is against standard Java best-practices.
Would it be possible to define an interface which declares all the methods in common between the PPL domains? The domains could implement this interface. This would allow to write generic code by just exploiting the subclass relationship.
Sorry for the long mail. I am very interested to know the opinion of the PPL developer and users on this points.
Cheers, --gianluca
On 02/25/2013 11:48 AM, Gianluca Amato wrote:
Hi all, I am using the PPL library from Java (hmm.. Scala, actually) to make experiments on abstract interpretation based static analysis. I would like to share some considerations about a couple of aspects of the Java bindings which I believe could be improved.
First aspect is performance. Obviously, using native libraries in the JVM always incurs a certain amount of performance penalty. I tried to develop a native implementation of Double_Box, and it is much faster than using the Double_Box fron PPL (the native implementation is not safe since I cannot control rounding mode within Java, but I do not think this impacts performance a lot). Obviously for more complex domains the overhead of calling native methods is proportionally lower.
My question is: is it possible to improve performance of the Java bindings? I am not an expert in JNI, but from a fast scan of the code, it seems the Java code passes to the native methods an object. The native code calls some accessor function to get the ptr field (which is the pointer to the native PPL object). It should be possible to rewrite the native functions so that they receive directly the pointer, and therefore do not need to call the accessor methods. From what I have heard, calling accessor methods is slow. Has this approach been pursued and abandoned or not? Do you think it could produce a noticeable improvement in performance?
Hi Gianluca.
I think that the first thing to do is to get a precise idea of where these inefficiencies are produced. For this we need some experimental evidence. If you provide us with a small testcase, we can try and help you in identifying the key spots.
The other point regards expressiveness. Java classes in PPL form a flat hierarchy. This makes difficult to write generic code, parametric w.r.t. the abstract domain. In C++ there is a flat hierarchy, too, but there templates may be used to write generic code effectively. In Java, the only choices to write generic code are either to use reflection (which is quite costly) or to write thin wrappers over the PPL classes, and put wrappers in a hierarchy, (which is tedious). Who works in Scala may use macros, but in any case, writing generic code is cumbersome, and the model of a flat hierarchy is against standard Java best-practices.
Would it be possible to define an interface which declares all the methods in common between the PPL domains? The domains could implement this interface. This would allow to write generic code by just exploiting the subclass relationship.
I guess it would be possible. The problem is that these similar looking interfaces are not exactly the same interface: some domains have different methods and a common interface would only be able to gather the common things, making the abstraction level less interesting. Also, as you said, it is going to be a tedious/repetitive work (this is also worsened by the fact that we are not using the Java bindings on a daily basis). This said, if you are willing to have a go on it and want to make your generic interface publicly available, we might later consider integrating it in the PPL (and maybe add any missing machinery to support such an integration).
Enea.
Sorry for the long mail. I am very interested to know the opinion of the PPL developer and users on this points.
Cheers, --gianluca
PPL-devel mailing list PPL-devel@cs.unipr.it http://www.cs.unipr.it/mailman/listinfo/ppl-devel
participants (2)
-
Enea Zaffanella -
Gianluca Amato