Incubator-mxnet: 1.3.0 release pre-built package contains a buggy openblas version. Causing gluon data loader with large num_workers to crash

[zhreshold]

Description

In latest mxnet pip prebuilt packages, when using gluon data loader with num_worker > certain number, it may raise error:

 OpenBLAS : Program will terminate because you tried to start too many threads.

Related issue: https://github.com/xianyi/OpenBLAS/issues/1735

Environment info (Required)

Ubuntu 16

pip install mxnet==1.3.0

from mxnet import gluon
dataset = xxx
batch_size = 16
loader = gluon.data.DataLoader(dataset, batch_size, num_workers=16)

[zhreshold]

@szha

[szha]

Looks like we just need to upgrade the openblas dependency to 0.3.3. Since it doesn't require code change in mxnet I will prepare a post-release to include the upgrade shortly.

[kalyc]

@mxnet-label-bot[Bug, BLAS]

[aaronmarkham]

@ankkhedia @Roshrini @lupesko - this may impact the windows instructions since we're currently having people use OpenBLAS v0.2.19 when building from source. WDYT @szha ?

[szha]

@aaronmarkham both versions should work. It's up to the user to decide which version of dependency to include.

[yajiedesign]

is impact the windows? Windows need to upgrade openblas?

[zhreshold]

i am only receiving report on ubunutu so far

[yajiedesign]

ok,not upgrade now,wait report.

[szha]

@zhreshold the 1.3.0.post0 releases contain the upgrade. Please verify.

[lupesko]

Thanks @szha.
@yajiedesign are you waiting for a confirmation before you publish the windows pip package for 1.3.0?

[zhreshold]

@szha Verified with same code, 1.3.0 failed and 1.3.0.post0 passed. Thanks for the timely fix!

[yajiedesign]

@lupesko At present, I have uploaded 1.3.0, waiting for report.