[ANN] Python compilers workshop at SciPy this year
Hi Nathaniel, I would like to suggest one more topic for the workshop. I see a big need for a library (jffi) similar to cffi but that provides a bridge to Java instead of C code. The ability to seamlessly work with native Java data/code would offer a huge improvement when python code needs to work with the Spark/Hadoop ecosystem. The current mechanisms which involve serializing data to/from Java can kill performance for some applications and can render Python unsuitable for these cases. John
Hi John I understand why you're bringing this up, but it's a huge project on it's own, worth at least a couple months worth of work. Without a dedicated effort from someone I'm worried it would not go anywhere. It's kind of separated from the other goals of the summit On Wed, Mar 23, 2016 at 8:16 PM, John Camara <john.m.camara@gmail.com> wrote:
Hi Fijal, I agree that jffi would be both a large project and without someone leading it, it would likely not get any where. But I tend to disagree that it would be a separate goal for the conference. I realize the goal of the summit is to talk about native-code compilation for Python and most would argue that means executing C code, assembly, or at the very least executing code at the speed of "C code". But the reality now is, numerical/scientific programming increasingly needs executing in a clustered environment. So I think we need to be careful to not only solve yesterday's problems but make sure we are covering the current day and future ones. Today, big data and analytics, which is driving most numerical/scientific programming, is becoming almost exclusively run in a clustered environment, with the Apache Spark ecosystem as the de facto standard. A few years back, Python's ace up its sleeve for the scientific community was the numpy/scipy ecosystem but we have recently lost that edge by falling behind in clustered computing. At this point in time our best move forward on the numerical/scientific fronts is to become best buddies with the Spark ecosystem and make sure we can bring bridge the numpy/scipy ecosystem to it. That is we merge the best of both worlds and suddenly Python becomes to go to language again for numerical/scientific computing. Of course we still need to address what should have been yesterday's problem and deal with the "native-code compilation" issues. John On Wed, Mar 23, 2016 at 2:47 PM, Maciej Fijalkowski <fijall@gmail.com> wrote:
On Mar 23, 2016 21:49, "Armin Rigo" <arigo@tunes.org> wrote:
need
There is also PyJNIus: https://pyjnius.readthedocs.org/en/latest/
Besides JPype and PyJNIus there is also https://www.py4j.org/. I haven't heard of JPype being used in any recent projects so I assuming it is outdated by now. PyJNIus gets used but I tend to only see it used on Android projects. The Py4J project gets used often in numerical/scientific projects mainly due to it use in PySpark. The problem with all these libraries is that they don't have a way to share large amounts of memory between the JVM and Python VMs and so large chunks of data have to be copied/serialized when going between the 2 VMs. Spark is the de facto standard in clustering computing at this point in time. At a high level Spark executes code that is distributed throughout a cluster so that the code being executed is as close as possible to where the data lives so as to minimize transferring of large amounts of data. The code that needs to be executed are packaged up into units called Resilient Distributed Dataset (RDD). RDDs are lazy evaluated and are essential graphs of the operations that need to be performed on the data. They are capable of reading data from many types of sources, outputting to multiple types of sources, containing the code that needs to be executed, and are also responsible to caching or keeping results in memory for future RDDs that maybe executed. If you write all your code in Java or Scala, its execution will be performed in JVMs distributed in the cluster. On the other hand, Spark does not limit its use to only Java based languages so Python can be used. In the case of Python the PySpark library is used. When Python is used, the PySpark library can be used to define the RDDs that will be executed under the JVM. In this scenario, only if required, the final results of the calculations will end up being passed to Python. I say only if necessary as its possible the end results may just be left in memory or to create an output such as an hdfs file in hadoop and does not need to be transferred to Python. Under this scenario the code is written in Python but effectively all the "real" work is performed under the JVM. Often someone writing Python is also going to want to perform some of the operations under Python. This can be done as the RDDs that are created can contain both operations that get performed under the JVM as well as Python (and of course other languages are supported). When Python is involved Spark will start up Python VMs on the required nodes so that the Python portions of the work can be performed. The Python VMs can either be CPython, PyPy or even a mix of both CPython and PyPy. The downside to using non Java languages is the overhead of passing data between the JVM and the Python VM as the memory is not shared between the processes but instead copied/serialized between them. Because this data is copied between the 2 VMs, anyone who writes Python code for this environment always has to be conscious of the data being copied between the processes so as to not let the amount of the extra overhead become a large burden. Quite often the goal will be to first perform the bulk of the operations under the JVM and then hopefully only a smaller subset of the data will have to be processed under Python. If this can be done then the overhead can be minimized and then there is essential no down sides to using Python in the pipeline of operations. If your unfortunate and need to perform some of the processing early in the pipline under Python and worse yet if there is a need to go back and forth many times between Python and Java the overhead of coping huge amounts of data can significantly slow things down which essentially puts Python at a disadvantage to Java. If it was possible to change the model of execution such that it was possible to embed the Python VM in the JVM or vice versa and that the memory could be shared between the 2 VMs the downside of using Python in this environment would be eliminated or at the very least minimized to the point where it is no longer an issue. Thus the need for a jffi library. There is a strong desire by many to use dynamic languages in these clustered environments and Python is likely in the best position to become the language of choice due to its ability to work with C based libraries and of course its syntax. The issues that hold Python back at this point is the serialization overhead, not so great state of packaging, and not having both the speed of the JIT and complete access to numpy/scipy ecosystem. Luckily for Python at this point there is no other dynamic language that is a clear winner today. But if too much time passes before these issues are solved I'm sure another language will step up to the plate. At this point my expectations is that Node could likely make a move. It already has the speed due to the Java Script JITs, it already has a great story for packaging and deployment, and its growth is exploding on the server side due to all the money being poured into it. What it strongly lacks today is the connection to C/legacy code, numerical/scientific modules and of course it also does not have a solution to the data copying overhead it also has with the JVM. Any way, this is just my 2 cents on what is currently holding Python back from taking off in this space. On Thu, Mar 24, 2016 at 2:32 AM, Hakan Ardo <hakan.ardo@gmail.com> wrote:
Hi John Thanks for explaining the current situation of the ecosystem. I'm not quite sure what your intention is. PyPy (and CPython) is very easy to embed through any C-level API, especially with the latest additions to cffi embedding. If someone feels like doing the work to share stuff that way (as I presume a lot of data presented in JVM can be represented as some pointer and shape how to access it), then he's obviously more than free to do so, I'm even willing to help with that. Now this seems like a medium-to-big size project that additionally will require quite a bit of community will to endorse. Are you willing to volunteer to work on such a project and dedicate a lot of time to it? If not, then there is no way you can convince us to volunteer our own time to do it - it's just too big and quite a bit far out of our usual areas of interest. If there is some commercial interest (and I think there might be) in pushing python and especially pypy further in that area, we might want to have a better story for numpy first, but then feel free to send those corporate interest people my way, we can maybe organize something. If you want us to do community service to push Python solutions in the area I have very little clue about however, I would like to politely decline. Cheers, fijal On Thu, Mar 24, 2016 at 2:22 PM, John Camara <john.m.camara@gmail.com> wrote:
Hi Fijal, I understand where your coming from and not trying to convince you to work on it. Just mainly trying to point out a need that may not be obvious to this community. I don't spend much time on big data and analytics so I don't have a lot of time to devote to this task. That could change in the future so you never know I may end up getting involved with this. At the end of the day I think it is the PSF, which needs to do an honest assessment of the current state of Python and in programming in general, so that they can help direct the future of Python. I think with an honest assessment it should be clear that it is absolutely necessary that a dynamic language have a JIT. Otherwise, a language like Node would not be growing so quickly on the server side. An honest assessment would conclude that Python needs to play a major role in big data and analytics as we don't want this to be another area where Python misses the boat. As with all languages other than JavaScript we missed playing an important role on web front end. More recently we missed out on mobile. I don't think it is good for us to miss out on big data. It would be a shame since we had such a strong scientific community which initially gave us a huge advantage over other communities. Missing out on big data might also be the driver that moves the scientific community in a different direction which would be a big loss to Python. I personally don't see any particular companies or industries that are willing to fund the tasks needed to solve these issues. It's not to say there are no more funds for Python projects its just likely no one company will be willing to fund these kinds of projects on their own. It really needs the PSF to coordinate these efforts but they seamed to be more focus on trying to make Python 3 a success instead of improving the overall health of the community. I believe that Python is in pretty good shape in being able to solve these issues but it just needs some funding and focus to get there. Hopefully the workshop will be successful and help create some focus. John On Thu, Mar 24, 2016 at 8:56 AM, Maciej Fijalkowski <fijall@gmail.com> wrote:
Hi John, On 24 March 2016 at 13:22, John Camara <john.m.camara@gmail.com> wrote:
(...) Thus the need for a jffi library.
When I hear "a jffi library" I'm thinking about a new library with a new API. I think what you would really like instead is to keep the existing libraries, but adapt them internally to allow tighter execution of the Python and Java VMs. I may be completely wrong about that, but you're also talking to the wrong guys in the first place :-) A bientôt, Armin.
Hi Armin, At a minimum tighter execution is required as well as sharing memory. But on the other hand you have raised the bar so high with cffi, having a clean and unbloated interface, that it would be nice if a library with a similar spirit existed for java. Having support in PyPy's JIT to remove all the marshalling types would be a big plus on top of the shared memory as well as some integration between the 2 GCs would likely be required. Maybe the best approach would be a combination of existing libraries and a new interface that allows for sharing of memory. Maybe similar to numpy arrays with a better API that avoids the pit falls of numpy relying on CPython semantics/implementation details. After all the only thing that needs to be eliminated is the copying/serialization of large data arrays/structures. John On Thu, Mar 24, 2016 at 12:20 PM, Armin Rigo <arigo@tunes.org> wrote:
It turns out there is some work in progress in the Spark project to share its memory with non JVM programs. See https://issues.apache.org/jira/browse/SPARK-10399. Once this is completed it should be fairly trivial to expose it to Python and then maybe JIT integration could be discussed at that time. This is a huge step forward over sharing Java objects. From the title of the ticket it appears it would be a c++ interface but looking at the pull request it looks like it will be a c interface. In the end the blocker may just come down to PyPy having complete support for Numpy. Without Numpy the success of this would be somewhat limited based on user expectations and without PyPy it maybe to slow for many applications. On Thu, Mar 24, 2016 at 1:11 PM, John Camara <john.m.camara@gmail.com> wrote:
Hi John I understand why you're bringing this up, but it's a huge project on it's own, worth at least a couple months worth of work. Without a dedicated effort from someone I'm worried it would not go anywhere. It's kind of separated from the other goals of the summit On Wed, Mar 23, 2016 at 8:16 PM, John Camara <john.m.camara@gmail.com> wrote:
Hi Fijal, I agree that jffi would be both a large project and without someone leading it, it would likely not get any where. But I tend to disagree that it would be a separate goal for the conference. I realize the goal of the summit is to talk about native-code compilation for Python and most would argue that means executing C code, assembly, or at the very least executing code at the speed of "C code". But the reality now is, numerical/scientific programming increasingly needs executing in a clustered environment. So I think we need to be careful to not only solve yesterday's problems but make sure we are covering the current day and future ones. Today, big data and analytics, which is driving most numerical/scientific programming, is becoming almost exclusively run in a clustered environment, with the Apache Spark ecosystem as the de facto standard. A few years back, Python's ace up its sleeve for the scientific community was the numpy/scipy ecosystem but we have recently lost that edge by falling behind in clustered computing. At this point in time our best move forward on the numerical/scientific fronts is to become best buddies with the Spark ecosystem and make sure we can bring bridge the numpy/scipy ecosystem to it. That is we merge the best of both worlds and suddenly Python becomes to go to language again for numerical/scientific computing. Of course we still need to address what should have been yesterday's problem and deal with the "native-code compilation" issues. John On Wed, Mar 23, 2016 at 2:47 PM, Maciej Fijalkowski <fijall@gmail.com> wrote:
On Mar 23, 2016 21:49, "Armin Rigo" <arigo@tunes.org> wrote:
need
There is also PyJNIus: https://pyjnius.readthedocs.org/en/latest/
Besides JPype and PyJNIus there is also https://www.py4j.org/. I haven't heard of JPype being used in any recent projects so I assuming it is outdated by now. PyJNIus gets used but I tend to only see it used on Android projects. The Py4J project gets used often in numerical/scientific projects mainly due to it use in PySpark. The problem with all these libraries is that they don't have a way to share large amounts of memory between the JVM and Python VMs and so large chunks of data have to be copied/serialized when going between the 2 VMs. Spark is the de facto standard in clustering computing at this point in time. At a high level Spark executes code that is distributed throughout a cluster so that the code being executed is as close as possible to where the data lives so as to minimize transferring of large amounts of data. The code that needs to be executed are packaged up into units called Resilient Distributed Dataset (RDD). RDDs are lazy evaluated and are essential graphs of the operations that need to be performed on the data. They are capable of reading data from many types of sources, outputting to multiple types of sources, containing the code that needs to be executed, and are also responsible to caching or keeping results in memory for future RDDs that maybe executed. If you write all your code in Java or Scala, its execution will be performed in JVMs distributed in the cluster. On the other hand, Spark does not limit its use to only Java based languages so Python can be used. In the case of Python the PySpark library is used. When Python is used, the PySpark library can be used to define the RDDs that will be executed under the JVM. In this scenario, only if required, the final results of the calculations will end up being passed to Python. I say only if necessary as its possible the end results may just be left in memory or to create an output such as an hdfs file in hadoop and does not need to be transferred to Python. Under this scenario the code is written in Python but effectively all the "real" work is performed under the JVM. Often someone writing Python is also going to want to perform some of the operations under Python. This can be done as the RDDs that are created can contain both operations that get performed under the JVM as well as Python (and of course other languages are supported). When Python is involved Spark will start up Python VMs on the required nodes so that the Python portions of the work can be performed. The Python VMs can either be CPython, PyPy or even a mix of both CPython and PyPy. The downside to using non Java languages is the overhead of passing data between the JVM and the Python VM as the memory is not shared between the processes but instead copied/serialized between them. Because this data is copied between the 2 VMs, anyone who writes Python code for this environment always has to be conscious of the data being copied between the processes so as to not let the amount of the extra overhead become a large burden. Quite often the goal will be to first perform the bulk of the operations under the JVM and then hopefully only a smaller subset of the data will have to be processed under Python. If this can be done then the overhead can be minimized and then there is essential no down sides to using Python in the pipeline of operations. If your unfortunate and need to perform some of the processing early in the pipline under Python and worse yet if there is a need to go back and forth many times between Python and Java the overhead of coping huge amounts of data can significantly slow things down which essentially puts Python at a disadvantage to Java. If it was possible to change the model of execution such that it was possible to embed the Python VM in the JVM or vice versa and that the memory could be shared between the 2 VMs the downside of using Python in this environment would be eliminated or at the very least minimized to the point where it is no longer an issue. Thus the need for a jffi library. There is a strong desire by many to use dynamic languages in these clustered environments and Python is likely in the best position to become the language of choice due to its ability to work with C based libraries and of course its syntax. The issues that hold Python back at this point is the serialization overhead, not so great state of packaging, and not having both the speed of the JIT and complete access to numpy/scipy ecosystem. Luckily for Python at this point there is no other dynamic language that is a clear winner today. But if too much time passes before these issues are solved I'm sure another language will step up to the plate. At this point my expectations is that Node could likely make a move. It already has the speed due to the Java Script JITs, it already has a great story for packaging and deployment, and its growth is exploding on the server side due to all the money being poured into it. What it strongly lacks today is the connection to C/legacy code, numerical/scientific modules and of course it also does not have a solution to the data copying overhead it also has with the JVM. Any way, this is just my 2 cents on what is currently holding Python back from taking off in this space. On Thu, Mar 24, 2016 at 2:32 AM, Hakan Ardo <hakan.ardo@gmail.com> wrote:
Hi John Thanks for explaining the current situation of the ecosystem. I'm not quite sure what your intention is. PyPy (and CPython) is very easy to embed through any C-level API, especially with the latest additions to cffi embedding. If someone feels like doing the work to share stuff that way (as I presume a lot of data presented in JVM can be represented as some pointer and shape how to access it), then he's obviously more than free to do so, I'm even willing to help with that. Now this seems like a medium-to-big size project that additionally will require quite a bit of community will to endorse. Are you willing to volunteer to work on such a project and dedicate a lot of time to it? If not, then there is no way you can convince us to volunteer our own time to do it - it's just too big and quite a bit far out of our usual areas of interest. If there is some commercial interest (and I think there might be) in pushing python and especially pypy further in that area, we might want to have a better story for numpy first, but then feel free to send those corporate interest people my way, we can maybe organize something. If you want us to do community service to push Python solutions in the area I have very little clue about however, I would like to politely decline. Cheers, fijal On Thu, Mar 24, 2016 at 2:22 PM, John Camara <john.m.camara@gmail.com> wrote:
Hi Fijal, I understand where your coming from and not trying to convince you to work on it. Just mainly trying to point out a need that may not be obvious to this community. I don't spend much time on big data and analytics so I don't have a lot of time to devote to this task. That could change in the future so you never know I may end up getting involved with this. At the end of the day I think it is the PSF, which needs to do an honest assessment of the current state of Python and in programming in general, so that they can help direct the future of Python. I think with an honest assessment it should be clear that it is absolutely necessary that a dynamic language have a JIT. Otherwise, a language like Node would not be growing so quickly on the server side. An honest assessment would conclude that Python needs to play a major role in big data and analytics as we don't want this to be another area where Python misses the boat. As with all languages other than JavaScript we missed playing an important role on web front end. More recently we missed out on mobile. I don't think it is good for us to miss out on big data. It would be a shame since we had such a strong scientific community which initially gave us a huge advantage over other communities. Missing out on big data might also be the driver that moves the scientific community in a different direction which would be a big loss to Python. I personally don't see any particular companies or industries that are willing to fund the tasks needed to solve these issues. It's not to say there are no more funds for Python projects its just likely no one company will be willing to fund these kinds of projects on their own. It really needs the PSF to coordinate these efforts but they seamed to be more focus on trying to make Python 3 a success instead of improving the overall health of the community. I believe that Python is in pretty good shape in being able to solve these issues but it just needs some funding and focus to get there. Hopefully the workshop will be successful and help create some focus. John On Thu, Mar 24, 2016 at 8:56 AM, Maciej Fijalkowski <fijall@gmail.com> wrote:
Hi John, On 24 March 2016 at 13:22, John Camara <john.m.camara@gmail.com> wrote:
(...) Thus the need for a jffi library.
When I hear "a jffi library" I'm thinking about a new library with a new API. I think what you would really like instead is to keep the existing libraries, but adapt them internally to allow tighter execution of the Python and Java VMs. I may be completely wrong about that, but you're also talking to the wrong guys in the first place :-) A bientôt, Armin.
Hi Armin, At a minimum tighter execution is required as well as sharing memory. But on the other hand you have raised the bar so high with cffi, having a clean and unbloated interface, that it would be nice if a library with a similar spirit existed for java. Having support in PyPy's JIT to remove all the marshalling types would be a big plus on top of the shared memory as well as some integration between the 2 GCs would likely be required. Maybe the best approach would be a combination of existing libraries and a new interface that allows for sharing of memory. Maybe similar to numpy arrays with a better API that avoids the pit falls of numpy relying on CPython semantics/implementation details. After all the only thing that needs to be eliminated is the copying/serialization of large data arrays/structures. John On Thu, Mar 24, 2016 at 12:20 PM, Armin Rigo <arigo@tunes.org> wrote:
It turns out there is some work in progress in the Spark project to share its memory with non JVM programs. See https://issues.apache.org/jira/browse/SPARK-10399. Once this is completed it should be fairly trivial to expose it to Python and then maybe JIT integration could be discussed at that time. This is a huge step forward over sharing Java objects. From the title of the ticket it appears it would be a c++ interface but looking at the pull request it looks like it will be a c interface. In the end the blocker may just come down to PyPy having complete support for Numpy. Without Numpy the success of this would be somewhat limited based on user expectations and without PyPy it maybe to slow for many applications. On Thu, Mar 24, 2016 at 1:11 PM, John Camara <john.m.camara@gmail.com> wrote:
participants (5)
-
Armin Rigo -
David Edelsohn -
Hakan Ardo -
John Camara -
Maciej Fijalkowski