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 <lau...@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
