mxnet-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Chris Olivier <cjolivie...@gmail.com>
Subject Re: Please remove conflicting Open MP version from CMake builds
Date Sat, 07 Dec 2019 15:12:11 GMT
btw trying to override a veto with a “lazy consensus” is not a valid
approach.

On Fri, Dec 6, 2019 at 8:44 PM Lausen, Leonard <lausen@amazon.com.invalid>
wrote:

> I think it's reasonable to assume that the Intel MKLDNN team is an
> "authorative"
> source about the issue of compilation with OpenMP and the OpenMP runtime
> library
> related issues. Thus I suggest we follow the recommendation of Intel
> MKLDNN team
> within the MXNet project.
>
> Looking through the Intel MKLDNN documentation, I find [1]:
>
> > DNNL uses OpenMP runtime library provided by the compiler.
>
> as well as
>
> > it's important to ensure that only one OpenMP runtime is used throughout
> the
> > application. Having more than one OpenMP runtime linked to an executable
> may
> > lead to undefined behavior including incorrect results or crashes.
>
> To keep our project maintainable and error free, I thus suggest we follow
> DNNL
> and use the OpenMP runtime library provided by the compiler.
> We have limited ressources and finding the root cause for any bugs
> resulting
> from linking multiple OpenMP libraries as currently done is, in my
> opinion. not
> a good use of time. We know it's due to undefined behavior and we know
> it's best
> practice to use OpenMP runtime library provided by the compiler. So let's
> just
> do that.
>
> I think given that MKL-DNN has also adopted the "OpenMP runtime library
> provided
> by the compiler" approach, this issue is not contentious anymore and
> qualifies
> for lazy consensus.
>
> Thus if there is no objection within 72 hours (lazy consensus), let's drop
> bundled LLVM OpenMP from master [2]. If we find any issues due to
> droppeing the
> bundled LLVM OpenMP, we can always add it back prior to the next release.
>
> Best regards
> Leonard
>
> [1]:
>
> https://github.com/intel/mkl-dnn/blob/433e086bf5d9e5ccfc9ec0b70322f931b6b1921d/doc/build/build_options.md#openmp
> (This is the updated reference from Anton's previous comment, based on the
> changes in MKLDNN done in the meantime
> https://github.com/apache/incubator-mxnet/pull/12160#issuecomment-415078066
> )
> [2]: Alike https://github.com/apache/incubator-mxnet/pull/12160
>
>
> On Fri, 2019-12-06 at 12:16 -0800, Pedro Larroy wrote:
> > I will try to stay on the sidelines for now since previous conversations
> > about OMP have not been productive here and I have spent way too much
> time
> > on this already, I'm not the first one giving up on trying to help with
> > this topic.
> >
> > I would be glad if you guys can work together and find a solution. I will
> > just put my understanding of the big picture hoping that it helps move it
> > forward.
> >
> >
> > Recently the intel omp library which seemed to have the best performance
> of
> > the 3 was removed from MKL.
> >
> > - There's 3 libraries in play, GNU Omp which is shipped with gcc (gomp),
> > LLVM openmp in 3rdparty (llvm-omp), Intel OMP when using MKL, which is
> > recently removed (iomp)
> >
> > - IOMP seems to have the best performance, there's stability issues
> > producing crashes sometimes but the impact seems relatively small for
> users
> > and developers. In general seems linking with a different OMP version
> that
> > the one shipped with the compiler is known to cause stability issues but
> > it's done anyway.
> >
> > - LLVM-OMP used when building with CMake, not used in the PIP releases or
> > when building with Make. Has stability issues, hangs when running in
> debug
> > mode during test execution and produces tons of assertions in debug mode.
> > Might have some small performance gains but there is no clear cut data
> that
> > showcases significant performance gains.
> >
> > - GOMP is the version shipped with GCC and the PIP wheels without MKL,
> has
> > no stability problems.
> >
> > As a ballpark, IOMP might give 10% performance improvement in some cases.
> >
> > We need to document well how users should tune and configure MXNet when
> > using OMP.
> >
> > As a developer, the safest bet is to use GOMP to be able to debug and
> > develop without issues. As a user of CPU inference / training you want to
> > run MKL so depends on how the Intel guys want to do things. My preference
> > as an engineer is always stability > speed.
> >
> > Related tickets:
> >
> > https://github.com/apache/incubator-mxnet/issues/16891
> >
> >
> https://github.com/apache/incubator-mxnet/issues/10856#issuecomment-562637931
> >
> >
> > https://github.com/apache/incubator-mxnet/issues/11417
> >
> > https://github.com/apache/incubator-mxnet/issues/15690
> >
> >
> >
> > On Fri, Dec 6, 2019 at 12:39 AM Lausen, Leonard
> <lausen@amazon.com.invalid>
> > wrote:
> >
> > > Is this related to
> https://github.com/apache/incubator-mxnet/issues/10856?
> > >
> > > I unlocked that Github issue based on the Apache Code of Conduct
> > > https://www.apache.org/foundation/policies/conduct#specific-guidelines
> > >
> > >
> > > On Sat, 2019-11-30 at 02:47 -0800, Pedro Larroy wrote:
> > > > (py3_venv) piotr@34-215-197-42:1:~/mxnet_1.6 (upstream_master)+$ ldd
> > > > build/libmxnet.so| grep -i openmp
> > > >         libomp.so =>
> > > > /home/piotr/mxnet_1.6/build/3rdparty/openmp/runtime/src/libomp.so
> > > > (0x00007fde0991d000)
> > > > (py3_venv) piotr@34-215-197-42:0:~/mxnet_1.6 (upstream_master)+$
> python
> > > > ~/deeplearning-benchmark/image_classification/infer_imagenet.py
> --use-rec
> > > > --batch-size 256 --dtype float32 --num-data-workers 40 --mode hybrid
> > > > --model resnet50_v2 --use-pretrained --kvstore local --log-interval 1
> > > > --rec-val ~/data/val-passthrough.rec --rec-val-idx
> > > > ~/data/val-passthrough.idx
> > > > INFO:root:Namespace(batch_norm=False, batch_size=256,
> > > > data_dir='~/.mxnet/datasets/imagenet', dataset_size=32,
> dtype='float32',
> > > > kvstore='local', last_gamma=False, log_interval=1,
> logging_dir='logs',
> > > > lr=0.1, lr_decay=0.1, lr_decay_epoch='40,60', lr_mode='step',
> > > > lr_poly_power=2, mode='hybrid', model='resnet50_v2', momentum=0.9,
> > > > num_epochs=3, num_gpus=0, num_workers=40,
> > > > rec_val='/home/piotr/data/val-passthrough.rec',
> > > > rec_val_idx='/home/piotr/data/val-passthrough.idx',
> save_dir='params',
> > > > save_frequency=0, top_k=0, use_pretrained=True, use_rec=True,
> > > use_se=False,
> > > > warmup_epochs=0, warmup_lr=0.0, wd=0.0001)
> > > > [10:42:02] ../src/io/iter_image_recordio_2.cc:178:
> ImageRecordIOParser2:
> > > > /home/piotr/data/val-passthrough.rec, use 36 threads for decoding..
> > > > INFO:root:Batch [0]
> > > > INFO:root:Top 1 accuracy: 0
> > > > INFO:root:warmup_throughput: 5 samples/sec warmup_time 43.150922
> > > > INFO:root:Batch [1]
> > > > INFO:root:Top 1 accuracy: 0
> > > > INFO:root:warmup_throughput: 6 samples/sec warmup_time 37.971927
> > > > INFO:root:Batch [2]
> > > > INFO:root:Top 1 accuracy: 0
> > > > INFO:root:warmup_throughput: 7 samples/sec warmup_time 35.755363
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > > (py3_venv) piotr@34-215-197-42:0:~/mxnet_1.6_plat_omp
> > > (upstream_master)+$
> > > > git st
> > > > On branch upstream_master
> > > > Your branch is up to date with 'origin/upstream_master'.
> > > >
> > > > Changes not staged for commit:
> > > >   (use "git add/rm <file>..." to update what will be committed)
> > > >   (use "git checkout -- <file>..." to discard changes in working
> > > directory)
> > > >         deleted:    3rdparty/openmp
> > > >
> > > > no changes added to commit (use "git add" and/or "git commit -a")
> > > > (py3_venv) piotr@34-215-197-42:1:~/mxnet_1.6_plat_omp
> > > (upstream_master)+$
> > > > ldd build/libmxnet.so | grep -i omp
> > > >         libgomp.so.1 => /usr/lib/x86_64-linux-gnu/libgomp.so.1
> > > > (0x00007f941241c000)
> > > >
> > > > (py3_venv) piotr@34-215-197-42:130:~/mxnet_1.6_plat_omp
> > > (upstream_master)+$
> > > > python
> ~/deeplearning-benchmark/image_classification/infer_imagenet.py
> > > > --use-rec --batch-size 256 --dtype float32 --num-data-workers 40
> --mode
> > > > hybrid --model resnet50_v2 --use-pretrained --kvstore local
> > > --log-interval
> > > > 1 --rec-val ~/data/val-passthrough.rec --rec-val-idx
> > > > ~/data/val-passthrough.idx
> > > > INFO:root:warmup_throughput: 147 samples/sec warmup_time 1.735117
> > > > INFO:root:Batch [16]
> > > > INFO:root:Top 1 accuracy: 0
> > > > INFO:root:warmup_throughput: 143 samples/sec warmup_time 1.785760
> > > > INFO:root:Batch [17]
> > > > INFO:root:Top 1 accuracy: 0
> > > > INFO:root:warmup_throughput: 148 samples/sec warmup_time 1.729033
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message