Re: Bottom up build
On 30.01.2014 23:10, Rob Weir wrote: On Wed, Jan 29, 2014 at 4:18 AM, Andre Fischer awf@gmail.com wrote: I would like to report some observations that I made when thinking about how to make building OpenOffice with one global makefile feasible. It will probably the last of build related mails in the near future. Traditional make uses a top-down approach. It starts with a target, 'all' by default, and looks at its dependencies. When one of those has to be made or is newer then the target then the target also has to be made. This is done recursively and depth-first. Every file on which 'all' has a direct or indirect dependency has to be checked. If we would build OpenOffice with one makefile (included makefiles don't count) then that are a lot of files to check. There are about 9800 cxx files, 3500 c files, 12000 hxx files, and lot of external headers. Checking the modification time of so many files is one of the reasons for the delay in , say, sw/ between starting make and its first action. As I don't have all global dependencies in a format that would allow experimation, I tried how long it would take to get the mtime (time of last modification) for all files, source, generated, compiled, linked, about 12. I wrote a small Java program for that. With a warm cache that takes about 23s. When run in 4 threads this reduced to less than 8s. Could be worse. But it also could be better because in general there are only a few files modified, usually files that you modified yourself in an editor. There is another approach, introduced, as far as I know, by the tup [1] build tool, that is bottom up. If you had something similar to the oracle of complexity theory, that gave you the list of modified files since the last build, you could find the depending files much faster. Faster for two reasons. Firstly, there is only one path in the dependency tree up towards the root (while there are many down from the root). Only targets on this path are affected by the modified file. Secondly, the dependency analysis is comparatively cheap. The expensive part is to determine the file modification times. If they where miraculously given then even the top-down approach would not take noticably longer. So I made another experiment to see if such an oracle can be created. Java 7 has the java.nio.file.WatchService that lets you monitor file modfifications in one directory. I registered it to all directories in our source tree (some 16000 directories). With the WatchService in place every file modification can be recorded and stored for later. On the next build you only have to check the set of modified files, not all files. Registering the directory watchers takes a couple of seconds but after that it does not cause any noticeable CPU activity. Any file modifications are reported almost at once. I do not have the framework in place to start a build with this information but I would expect it to be as fast as compiling the modified files and linking takes. The tup website references a paper [2] in which the established top-down approaches are called alpha alogithms and the new bottom-up approach is called beta algorithm. Tup has implemented a file modification watcher (in C or C++) only for Linux. On Windows it just scans all files (for which it needs a little more time than my Java program, maybe it does not use more than one thread). This is something that we should keep in mind for when we ever should get a build solution with global dependencies and this build tool would turn out to be too slow. If can find the source code of my Java experiments at [3]. If nothing else you can see an application of the ForkJoinPool that allowed my to write the parallel file system scan in just a few lines. There is also an alternative implementation that uses the ExecutorService (with a fixed thread pool) which needs a few more lines of code. And there is of course the use of the WatchService. Has anyone read this book? http://www.amazon.com/Large-Scale-Software-Design-John-Lakos/dp/0201633620 It was on my list to read for many years. From what I've seen it suggests design approaches to the improve build times. So things that go beyond what you can do by just changing build files, more fundamental changes to how interfaces are defined. I have not but I might, thanks for the hint. I agree that improving our software design is always a good idea but I would not change our design just to make the build faster. Many of our code related problems are caused by the design and the limitations of the build system. Examples are the individual building of directories (in dmake modules) and creation of one library per directory or by the many ugly tricks that avoid becoming incompatible (the need to compile files in other modules). Otherwise I wonder if we're trying to optimize a bubble sort? I would smile about this if I had not seen a handcrafted sort algorithm in our build scripts that
Re: Bottom up build
On Wed, Jan 29, 2014 at 1:54 AM, jan i j...@apache.org wrote: On 29 January 2014 10:18, Andre Fischer awf@gmail.com wrote: I would like to report some observations that I made when thinking about how to make building OpenOffice with one global makefile feasible. It will probably the last of build related mails in the near future. Traditional make uses a top-down approach. It starts with a target, 'all' by default, and looks at its dependencies. When one of those has to be made or is newer then the target then the target also has to be made. This is done recursively and depth-first. Every file on which 'all' has a direct or indirect dependency has to be checked. If we would build OpenOffice with one makefile (included makefiles don't count) then that are a lot of files to check. There are about 9800 cxx files, 3500 c files, 12000 hxx files, and lot of external headers. Checking the modification time of so many files is one of the reasons for the delay in , say, sw/ between starting make and its first action. As I don't have all global dependencies in a format that would allow experimation, I tried how long it would take to get the mtime (time of last modification) for all files, source, generated, compiled, linked, about 12. I wrote a small Java program for that. With a warm cache that takes about 23s. When run in 4 threads this reduced to less than 8s. Could be worse. But it also could be better because in general there are only a few files modified, usually files that you modified yourself in an editor. There is another approach, introduced, as far as I know, by the tup [1] build tool, that is bottom up. If you had something similar to the oracle of complexity theory, that gave you the list of modified files since the last build, you could find the depending files much faster. Faster for two reasons. Firstly, there is only one path in the dependency tree up towards the root (while there are many down from the root). Only targets on this path are affected by the modified file. Secondly, the dependency analysis is comparatively cheap. The expensive part is to determine the file modification times. If they where miraculously given then even the top-down approach would not take noticably longer. So I made another experiment to see if such an oracle can be created. Java 7 has the java.nio.file.WatchService that lets you monitor file modfifications in one directory. I registered it to all directories in our source tree (some 16000 directories). With the WatchService in place every file modification can be recorded and stored for later. On the next build you only have to check the set of modified files, not all files. Registering the directory watchers takes a couple of seconds but after that it does not cause any noticeable CPU activity. Any file modifications are reported almost at once. I do not have the framework in place to start a build with this information but I would expect it to be as fast as compiling the modified files and linking takes. The tup website references a paper [2] in which the established top-down approaches are called alpha alogithms and the new bottom-up approach is called beta algorithm. Tup has implemented a file modification watcher (in C or C++) only for Linux. On Windows it just scans all files (for which it needs a little more time than my Java program, maybe it does not use more than one thread). This is something that we should keep in mind for when we ever should get a build solution with global dependencies and this build tool would turn out to be too slow. If can find the source code of my Java experiments at [3]. If nothing else you can see an application of the ForkJoinPool that allowed my to write the parallel file system scan in just a few lines. There is also an alternative implementation that uses the ExecutorService (with a fixed thread pool) which needs a few more lines of code. And there is of course the use of the WatchService. It's really interesting to read these observations and test cases... we have a large and complicated source tree and just seeing what can be observed about it is fascinating to me. Thanks for writing down your observations which I find highly interesting. I hope your stop on writing about build does not include giving your opinion on my ideas in the future as well. For the record the capstone project, and my little hobby project Build R.I.P. follow a third idea: We have a clear seperation of module build and central (total) build. +1 I would certainly go for this. A while back someone asked about ye olde ld approach -- all modules compiled/built and then linked later down the road. If we could somehow do something to get back to that idea in a more friendly modern way, it would certainly make working on specific areas more feasible.
Re: Bottom up build
On Wed, Jan 29, 2014 at 4:18 AM, Andre Fischer awf@gmail.com wrote: I would like to report some observations that I made when thinking about how to make building OpenOffice with one global makefile feasible. It will probably the last of build related mails in the near future. Traditional make uses a top-down approach. It starts with a target, 'all' by default, and looks at its dependencies. When one of those has to be made or is newer then the target then the target also has to be made. This is done recursively and depth-first. Every file on which 'all' has a direct or indirect dependency has to be checked. If we would build OpenOffice with one makefile (included makefiles don't count) then that are a lot of files to check. There are about 9800 cxx files, 3500 c files, 12000 hxx files, and lot of external headers. Checking the modification time of so many files is one of the reasons for the delay in , say, sw/ between starting make and its first action. As I don't have all global dependencies in a format that would allow experimation, I tried how long it would take to get the mtime (time of last modification) for all files, source, generated, compiled, linked, about 12. I wrote a small Java program for that. With a warm cache that takes about 23s. When run in 4 threads this reduced to less than 8s. Could be worse. But it also could be better because in general there are only a few files modified, usually files that you modified yourself in an editor. There is another approach, introduced, as far as I know, by the tup [1] build tool, that is bottom up. If you had something similar to the oracle of complexity theory, that gave you the list of modified files since the last build, you could find the depending files much faster. Faster for two reasons. Firstly, there is only one path in the dependency tree up towards the root (while there are many down from the root). Only targets on this path are affected by the modified file. Secondly, the dependency analysis is comparatively cheap. The expensive part is to determine the file modification times. If they where miraculously given then even the top-down approach would not take noticably longer. So I made another experiment to see if such an oracle can be created. Java 7 has the java.nio.file.WatchService that lets you monitor file modfifications in one directory. I registered it to all directories in our source tree (some 16000 directories). With the WatchService in place every file modification can be recorded and stored for later. On the next build you only have to check the set of modified files, not all files. Registering the directory watchers takes a couple of seconds but after that it does not cause any noticeable CPU activity. Any file modifications are reported almost at once. I do not have the framework in place to start a build with this information but I would expect it to be as fast as compiling the modified files and linking takes. The tup website references a paper [2] in which the established top-down approaches are called alpha alogithms and the new bottom-up approach is called beta algorithm. Tup has implemented a file modification watcher (in C or C++) only for Linux. On Windows it just scans all files (for which it needs a little more time than my Java program, maybe it does not use more than one thread). This is something that we should keep in mind for when we ever should get a build solution with global dependencies and this build tool would turn out to be too slow. If can find the source code of my Java experiments at [3]. If nothing else you can see an application of the ForkJoinPool that allowed my to write the parallel file system scan in just a few lines. There is also an alternative implementation that uses the ExecutorService (with a fixed thread pool) which needs a few more lines of code. And there is of course the use of the WatchService. Has anyone read this book? http://www.amazon.com/Large-Scale-Software-Design-John-Lakos/dp/0201633620 It was on my list to read for many years. From what I've seen it suggests design approaches to the improve build times. So things that go beyond what you can do by just changing build files, more fundamental changes to how interfaces are defined. Otherwise I wonder if we're trying to optimize a bubble sort? -Rob Regards, Andre [1] http://gittup.org/tup/ [2] http://gittup.org/tup/build_system_rules_and_algorithms.pdf [3] http://people.apache.org/~af/test.zip - To unsubscribe, e-mail: dev-unsubscr...@openoffice.apache.org For additional commands, e-mail: dev-h...@openoffice.apache.org - To unsubscribe, e-mail: dev-unsubscr...@openoffice.apache.org For additional commands, e-mail: dev-h...@openoffice.apache.org
Re: Bottom up build
On 30 January 2014 23:10, Rob Weir robw...@apache.org wrote: On Wed, Jan 29, 2014 at 4:18 AM, Andre Fischer awf@gmail.com wrote: I would like to report some observations that I made when thinking about how to make building OpenOffice with one global makefile feasible. It will probably the last of build related mails in the near future. Traditional make uses a top-down approach. It starts with a target, 'all' by default, and looks at its dependencies. When one of those has to be made or is newer then the target then the target also has to be made. This is done recursively and depth-first. Every file on which 'all' has a direct or indirect dependency has to be checked. If we would build OpenOffice with one makefile (included makefiles don't count) then that are a lot of files to check. There are about 9800 cxx files, 3500 c files, 12000 hxx files, and lot of external headers. Checking the modification time of so many files is one of the reasons for the delay in , say, sw/ between starting make and its first action. As I don't have all global dependencies in a format that would allow experimation, I tried how long it would take to get the mtime (time of last modification) for all files, source, generated, compiled, linked, about 12. I wrote a small Java program for that. With a warm cache that takes about 23s. When run in 4 threads this reduced to less than 8s. Could be worse. But it also could be better because in general there are only a few files modified, usually files that you modified yourself in an editor. There is another approach, introduced, as far as I know, by the tup [1] build tool, that is bottom up. If you had something similar to the oracle of complexity theory, that gave you the list of modified files since the last build, you could find the depending files much faster. Faster for two reasons. Firstly, there is only one path in the dependency tree up towards the root (while there are many down from the root). Only targets on this path are affected by the modified file. Secondly, the dependency analysis is comparatively cheap. The expensive part is to determine the file modification times. If they where miraculously given then even the top-down approach would not take noticably longer. So I made another experiment to see if such an oracle can be created. Java 7 has the java.nio.file.WatchService that lets you monitor file modfifications in one directory. I registered it to all directories in our source tree (some 16000 directories). With the WatchService in place every file modification can be recorded and stored for later. On the next build you only have to check the set of modified files, not all files. Registering the directory watchers takes a couple of seconds but after that it does not cause any noticeable CPU activity. Any file modifications are reported almost at once. I do not have the framework in place to start a build with this information but I would expect it to be as fast as compiling the modified files and linking takes. The tup website references a paper [2] in which the established top-down approaches are called alpha alogithms and the new bottom-up approach is called beta algorithm. Tup has implemented a file modification watcher (in C or C++) only for Linux. On Windows it just scans all files (for which it needs a little more time than my Java program, maybe it does not use more than one thread). This is something that we should keep in mind for when we ever should get a build solution with global dependencies and this build tool would turn out to be too slow. If can find the source code of my Java experiments at [3]. If nothing else you can see an application of the ForkJoinPool that allowed my to write the parallel file system scan in just a few lines. There is also an alternative implementation that uses the ExecutorService (with a fixed thread pool) which needs a few more lines of code. And there is of course the use of the WatchService. Has anyone read this book? http://www.amazon.com/Large-Scale-Software-Design-John-Lakos/dp/0201633620 It was on my list to read for many years. From what I've seen it suggests design approaches to the improve build times. So things that go beyond what you can do by just changing build files, more fundamental changes to how interfaces are defined. Have read it, the book goes more into C++ structures and design, than the actual build process. If you have a pure C++ project, you can do a lot of speed improvement by definining the classes for speed instead of purity. Its quite a good book, but have very little for the AOO build system. Otherwise I wonder if we're trying to optimize a bubble sort? No we are trying to moving away from 3-4 build components trying to do the same thing and each sub-optimized. In other words, our
Bottom up build
I would like to report some observations that I made when thinking about how to make building OpenOffice with one global makefile feasible. It will probably the last of build related mails in the near future. Traditional make uses a top-down approach. It starts with a target, 'all' by default, and looks at its dependencies. When one of those has to be made or is newer then the target then the target also has to be made. This is done recursively and depth-first. Every file on which 'all' has a direct or indirect dependency has to be checked. If we would build OpenOffice with one makefile (included makefiles don't count) then that are a lot of files to check. There are about 9800 cxx files, 3500 c files, 12000 hxx files, and lot of external headers. Checking the modification time of so many files is one of the reasons for the delay in , say, sw/ between starting make and its first action. As I don't have all global dependencies in a format that would allow experimation, I tried how long it would take to get the mtime (time of last modification) for all files, source, generated, compiled, linked, about 12. I wrote a small Java program for that. With a warm cache that takes about 23s. When run in 4 threads this reduced to less than 8s. Could be worse. But it also could be better because in general there are only a few files modified, usually files that you modified yourself in an editor. There is another approach, introduced, as far as I know, by the tup [1] build tool, that is bottom up. If you had something similar to the oracle of complexity theory, that gave you the list of modified files since the last build, you could find the depending files much faster. Faster for two reasons. Firstly, there is only one path in the dependency tree up towards the root (while there are many down from the root). Only targets on this path are affected by the modified file. Secondly, the dependency analysis is comparatively cheap. The expensive part is to determine the file modification times. If they where miraculously given then even the top-down approach would not take noticably longer. So I made another experiment to see if such an oracle can be created. Java 7 has the java.nio.file.WatchService that lets you monitor file modfifications in one directory. I registered it to all directories in our source tree (some 16000 directories). With the WatchService in place every file modification can be recorded and stored for later. On the next build you only have to check the set of modified files, not all files. Registering the directory watchers takes a couple of seconds but after that it does not cause any noticeable CPU activity. Any file modifications are reported almost at once. I do not have the framework in place to start a build with this information but I would expect it to be as fast as compiling the modified files and linking takes. The tup website references a paper [2] in which the established top-down approaches are called alpha alogithms and the new bottom-up approach is called beta algorithm. Tup has implemented a file modification watcher (in C or C++) only for Linux. On Windows it just scans all files (for which it needs a little more time than my Java program, maybe it does not use more than one thread). This is something that we should keep in mind for when we ever should get a build solution with global dependencies and this build tool would turn out to be too slow. If can find the source code of my Java experiments at [3]. If nothing else you can see an application of the ForkJoinPool that allowed my to write the parallel file system scan in just a few lines. There is also an alternative implementation that uses the ExecutorService (with a fixed thread pool) which needs a few more lines of code. And there is of course the use of the WatchService. Regards, Andre [1] http://gittup.org/tup/ [2] http://gittup.org/tup/build_system_rules_and_algorithms.pdf [3] http://people.apache.org/~af/test.zip - To unsubscribe, e-mail: dev-unsubscr...@openoffice.apache.org For additional commands, e-mail: dev-h...@openoffice.apache.org
Re: Bottom up build
On 29 January 2014 10:18, Andre Fischer awf@gmail.com wrote: I would like to report some observations that I made when thinking about how to make building OpenOffice with one global makefile feasible. It will probably the last of build related mails in the near future. Traditional make uses a top-down approach. It starts with a target, 'all' by default, and looks at its dependencies. When one of those has to be made or is newer then the target then the target also has to be made. This is done recursively and depth-first. Every file on which 'all' has a direct or indirect dependency has to be checked. If we would build OpenOffice with one makefile (included makefiles don't count) then that are a lot of files to check. There are about 9800 cxx files, 3500 c files, 12000 hxx files, and lot of external headers. Checking the modification time of so many files is one of the reasons for the delay in , say, sw/ between starting make and its first action. As I don't have all global dependencies in a format that would allow experimation, I tried how long it would take to get the mtime (time of last modification) for all files, source, generated, compiled, linked, about 12. I wrote a small Java program for that. With a warm cache that takes about 23s. When run in 4 threads this reduced to less than 8s. Could be worse. But it also could be better because in general there are only a few files modified, usually files that you modified yourself in an editor. There is another approach, introduced, as far as I know, by the tup [1] build tool, that is bottom up. If you had something similar to the oracle of complexity theory, that gave you the list of modified files since the last build, you could find the depending files much faster. Faster for two reasons. Firstly, there is only one path in the dependency tree up towards the root (while there are many down from the root). Only targets on this path are affected by the modified file. Secondly, the dependency analysis is comparatively cheap. The expensive part is to determine the file modification times. If they where miraculously given then even the top-down approach would not take noticably longer. So I made another experiment to see if such an oracle can be created. Java 7 has the java.nio.file.WatchService that lets you monitor file modfifications in one directory. I registered it to all directories in our source tree (some 16000 directories). With the WatchService in place every file modification can be recorded and stored for later. On the next build you only have to check the set of modified files, not all files. Registering the directory watchers takes a couple of seconds but after that it does not cause any noticeable CPU activity. Any file modifications are reported almost at once. I do not have the framework in place to start a build with this information but I would expect it to be as fast as compiling the modified files and linking takes. The tup website references a paper [2] in which the established top-down approaches are called alpha alogithms and the new bottom-up approach is called beta algorithm. Tup has implemented a file modification watcher (in C or C++) only for Linux. On Windows it just scans all files (for which it needs a little more time than my Java program, maybe it does not use more than one thread). This is something that we should keep in mind for when we ever should get a build solution with global dependencies and this build tool would turn out to be too slow. If can find the source code of my Java experiments at [3]. If nothing else you can see an application of the ForkJoinPool that allowed my to write the parallel file system scan in just a few lines. There is also an alternative implementation that uses the ExecutorService (with a fixed thread pool) which needs a few more lines of code. And there is of course the use of the WatchService. Thanks for writing down your observations which I find highly interesting. I hope your stop on writing about build does not include giving your opinion on my ideas in the future as well. For the record the capstone project, and my little hobby project Build R.I.P. follow a third idea: We have a clear seperation of module build and central (total) build. The module makefile knows how to build the module, and the central makefile knows the relation between modules. The makefile in each module touched a file, and the central makefile only controls that file. But youir idea of watching for changes is very interesting. rgds jan I. Andre [1] http://gittup.org/tup/ [2] http://gittup.org/tup/build_system_rules_and_algorithms.pdf [3] http://people.apache.org/~af/test.zip - To unsubscribe, e-mail: dev-unsubscr...@openoffice.apache.org For additional commands, e-mail: dev-h...@openoffice.apache.org
