Re: [DISCUSS] Current Publish problems
Good stuff Zach and Qing! Great feedback Edison - would be great if you leave it in the wiki, so that it is saved in the context of the doc with other feedback, such as Kellen's. On Mon, Jan 28, 2019 at 1:58 AM edisongust...@gmail.com < edisongust...@gmail.com> wrote: > Hello all, > > First let me introduce myself: > > My name is Edison Gustavo Muenz. I have worked most of my career with C++, > Windows and Linux. I am a big fan of machine learning and now I joined > Amazon in Berlin to work on MXNet. > > I would like to give some comments on the document posted: > > # change publish OS (Severe) > > As a rule of thumb, when providing your own binaries on linux, we should > always try to compile with oldest glibc possible. Using CentOS7 for this > regard (if possible due to the CUDA issues) is the way to go. > > # Using Cent OS 7 > > > However, all of the current GPU build scripts would be unavailable since > nvidia does not provide the corresponding packages for rpm. In this case, > we may need to go with NVIDIA Docker for Cent OS 7 and that only provide a > limited versions of CUDA. > > > List of CUDA that NVIDIA supporting for Cent OS 7: > > CUDA 10, 9.2, 9.1, 9.0, 8.0, 7.5 > > From what I saw in the link provided ( > https://hub.docker.com/r/nvidia/cuda/), this list of versions is even > bigger than the list of versions supported on Ubuntu 16.04. > > What am I missing? > > > Another problem we may see is the performance and stability difference > on the backend we built since we downgrade libc from 2.19 to 2.17 > > I would like to first give a brief intro so that we're all on the same > page. If you already know how libc versioning works, then you can skip this > part > > ## Brief intro on how libc versioning works > > In libc each symbol provided by libc has 2 components: > - symbol name > - version > > This can be seen with: > > ``` > $ objdump -T /lib/x86_64-linux-gnu/libc.so.6 | grep memcpy > 000bd4a0 w DF .text 0009 GLIBC_2.2.5 wmemcpy > 001332f0 gDF .text 0019 GLIBC_2.4 __wmemcpy_chk > 0009f0e0 g iD .text 00ca GLIBC_2.14 memcpy > 000bb460 gDF .text 0028 (GLIBC_2.2.5) memcpy > 001318a0 g iD .text 00ca GLIBC_2.3.4 __memcpy_chk > ``` > > So it can be seen that there are different memory addresses for each > version of memcpy. > > When linking a binary, the linker will always choose the most recent > version of the libc symbol. > > An example: > - your program uses the `memcpy` symbol > - when linking, the linker will choose `memcpy` at version 2.14 > (latest) > > When executing the binary then the libc provided on your system must have > a memcpy at version 2.14, otherwise you get the following error: > > /lib/x86_64-linux-gnu/libm.so.6: version `libc_2.23' not found > (required by /tmp/mxnet6145590735071079280/libmxnet.so) > > Also, a symbol has its version increased when there are breaking changes. > So, libc will only increase the version of a symbol if any of its > inputs/outputs changed in a non-compatible way (eg.: Changing the type of a > field to a non-compatible type, like int -> short). > > ## Performance difference between versions 2.17 and 2.19 > > This website is really handy for this: > https://abi-laboratory.pro/?view=timeline=glibc > > If we look at the links: > > - > https://abi-laboratory.pro/index.php?view=objects_report=glibc=2.18=2.19 > - > https://abi-laboratory.pro/index.php?view=objects_report=glibc=2.17=2.18 > > You can see that their binary compatibility is fine since no significant > changes were made between these versions that could compromise the > performance. > > Finally, I want to thank everyone for letting me part of this community. > > On 2019/01/23 21:48:48, kellen sunderland > wrote: > > Hey Qing, thanks for the summary and to everyone for automating the > > deployment process. I've left a few comments on the doc. > > > > On Wed, Jan 23, 2019 at 11:46 AM Qing Lan wrote: > > > > > Hi all, > > > > > > Recently Zach announced the availability for MXNet Maven publishing > > > pipeline and general static-build instructions. In order to make it > better, > > > I drafted a document that includes the problems we have for this > pipeline: > > > > https://cwiki.apache.org/confluence/display/MXNET/Outstanding+problems+with+publishing > . > > > Some of them may need to be addressed very soon. > > > > > > Please kindly review and leave any comments you may have in this > thread or > > > in the document. > > > > > > thanks, > > > Qing > > > > > > > > >
Re: [DISCUSS] Current Publish problems
Hello all, First let me introduce myself: My name is Edison Gustavo Muenz. I have worked most of my career with C++, Windows and Linux. I am a big fan of machine learning and now I joined Amazon in Berlin to work on MXNet. I would like to give some comments on the document posted: # change publish OS (Severe) As a rule of thumb, when providing your own binaries on linux, we should always try to compile with oldest glibc possible. Using CentOS7 for this regard (if possible due to the CUDA issues) is the way to go. # Using Cent OS 7 > However, all of the current GPU build scripts would be unavailable since > nvidia does not provide the corresponding packages for rpm. In this case, we > may need to go with NVIDIA Docker for Cent OS 7 and that only provide a > limited versions of CUDA. > List of CUDA that NVIDIA supporting for Cent OS 7: > CUDA 10, 9.2, 9.1, 9.0, 8.0, 7.5 >From what I saw in the link provided (https://hub.docker.com/r/nvidia/cuda/), >this list of versions is even bigger than the list of versions supported on >Ubuntu 16.04. What am I missing? > Another problem we may see is the performance and stability difference on the > backend we built since we downgrade libc from 2.19 to 2.17 I would like to first give a brief intro so that we're all on the same page. If you already know how libc versioning works, then you can skip this part ## Brief intro on how libc versioning works In libc each symbol provided by libc has 2 components: - symbol name - version This can be seen with: ``` $ objdump -T /lib/x86_64-linux-gnu/libc.so.6 | grep memcpy 000bd4a0 w DF .text 0009 GLIBC_2.2.5 wmemcpy 001332f0 gDF .text 0019 GLIBC_2.4 __wmemcpy_chk 0009f0e0 g iD .text 00ca GLIBC_2.14 memcpy 000bb460 gDF .text 0028 (GLIBC_2.2.5) memcpy 001318a0 g iD .text 00ca GLIBC_2.3.4 __memcpy_chk ``` So it can be seen that there are different memory addresses for each version of memcpy. When linking a binary, the linker will always choose the most recent version of the libc symbol. An example: - your program uses the `memcpy` symbol - when linking, the linker will choose `memcpy` at version 2.14 (latest) When executing the binary then the libc provided on your system must have a memcpy at version 2.14, otherwise you get the following error: /lib/x86_64-linux-gnu/libm.so.6: version `libc_2.23' not found (required by /tmp/mxnet6145590735071079280/libmxnet.so) Also, a symbol has its version increased when there are breaking changes. So, libc will only increase the version of a symbol if any of its inputs/outputs changed in a non-compatible way (eg.: Changing the type of a field to a non-compatible type, like int -> short). ## Performance difference between versions 2.17 and 2.19 This website is really handy for this: https://abi-laboratory.pro/?view=timeline=glibc If we look at the links: - https://abi-laboratory.pro/index.php?view=objects_report=glibc=2.18=2.19 - https://abi-laboratory.pro/index.php?view=objects_report=glibc=2.17=2.18 You can see that their binary compatibility is fine since no significant changes were made between these versions that could compromise the performance. Finally, I want to thank everyone for letting me part of this community. On 2019/01/23 21:48:48, kellen sunderland wrote: > Hey Qing, thanks for the summary and to everyone for automating the > deployment process. I've left a few comments on the doc. > > On Wed, Jan 23, 2019 at 11:46 AM Qing Lan wrote: > > > Hi all, > > > > Recently Zach announced the availability for MXNet Maven publishing > > pipeline and general static-build instructions. In order to make it better, > > I drafted a document that includes the problems we have for this pipeline: > > https://cwiki.apache.org/confluence/display/MXNET/Outstanding+problems+with+publishing. > > Some of them may need to be addressed very soon. > > > > Please kindly review and leave any comments you may have in this thread or > > in the document. > > > > thanks, > > Qing > > > > >
Re: [DISCUSS] Current Publish problems
Hey Qing, thanks for the summary and to everyone for automating the deployment process. I've left a few comments on the doc. On Wed, Jan 23, 2019 at 11:46 AM Qing Lan wrote: > Hi all, > > Recently Zach announced the availability for MXNet Maven publishing > pipeline and general static-build instructions. In order to make it better, > I drafted a document that includes the problems we have for this pipeline: > https://cwiki.apache.org/confluence/display/MXNET/Outstanding+problems+with+publishing. > Some of them may need to be addressed very soon. > > Please kindly review and leave any comments you may have in this thread or > in the document. > > thanks, > Qing > >
[DISCUSS] Current Publish problems
Hi all, Recently Zach announced the availability for MXNet Maven publishing pipeline and general static-build instructions. In order to make it better, I drafted a document that includes the problems we have for this pipeline: https://cwiki.apache.org/confluence/display/MXNET/Outstanding+problems+with+publishing. Some of them may need to be addressed very soon. Please kindly review and leave any comments you may have in this thread or in the document. thanks, Qing