Incubator-mxnet: [RFC] Large MXNet source files causing CI build failures

[mseth10]

Problem statement

MXNet CI is running OOM [1] while building MXNet binaries for unix-cpu and unix-gpu stages. This is an intermittent failure and the work around is to re-trigger CI a few times. The issue is caused due to some of the numpy .cc files being too large causing gcc to use too much memory. The issue was not pronounced with gcc7, but with the recent update to use gcc8 [2] for CI builds, we have started to see this OOM error.

The fix is to refactor the numpy .cc files into smaller files so that the objects created during compilation don't use much memory. Here is the list of the largest objects (>10MB in size) generated currently on Mac CPU build:

 11M    ./operator/numpy/linalg/np_norm_backward.cc.o
 11M    ./operator/numpy/np_kron.cc.o
 11M    ./operator/numpy/random/np_location_scale_op.cc.o
 12M    ./operator/numpy/np_insert_op_slice.cc.o
 12M    ./operator/numpy/np_insert_op_tensor.cc.o
 13M    ./operator/numpy/np_elemwise_broadcast_op_extended_sec.cc.o
 13M    ./operator/numpy/np_elemwise_unary_op_basic.cc.o
 13M    ./operator/numpy/np_percentile_op.cc.o
 14M    ./operator/numpy/np_matrix_op.cc.o
 14M    ./operator/numpy/np_moments_op.cc.o
 14M    ./operator/numpy/np_where_op.cc.o
 15M    ./operator/numpy/np_einsum_op.cc.o
 16M    ./operator/numpy/np_elemwise_broadcast_op_extended.cc.o
 21M    ./operator/numpy/np_broadcast_reduce_op_value.cc.o
 22M    ./operator/numpy/linalg/np_norm_forward.cc.o
 24M    ./operator/numpy/np_elemwise_broadcast_op.cc.o
 34M    ./operator/numpy/np_elemwise_broadcast_logic_op.cc.o

The corresponding cc files to above objects contains more than 210 operator registrations, and to refactor those into smaller files will need a considerable time and effort from the community. With 5 operators per day, that's more than 40 days of developers effort.

Proposed solutions

Option 1: We keep using gcc8 for CI builds and start working on refactoring these numpy .cc files. This would mean the community will have to face the CI failures for 40 days (could be less if more community members contribute).

Option 2: We go back to using gcc7 for CI builds, potentially solving the CI problem immediately, while we work on refactoring the numpy files. Reverting to gcc7 would take 2 days and then refactoring would take another 40 days.

I personally would prefer Option 2 for the reason that it saves contributors time in getting their PRs merged quickly, as well as saves on the CI resources. Would like to request community feedback on the same.

Going forward we also need to add a check to MXNet CI for build time memory usage. Any ideas for the same would be highly appreciated.

References

• [1] CI failure: https://jenkins.mxnet-ci.amazon-ml.com/blue/organizations/jenkins/mxnet-validation%2Funix-cpu/detail/PR-19588/12/pipeline/
• [2] CI builds gcc update commit: https://github.com/apache/incubator-mxnet/commit/afc76b0f82839dfe00f7cab04be6de3df94564fe

[wkcn]

There is a related issue: https://github.com/apache/incubator-mxnet/issues/18501
I also prefer Option 2 to solve the CI problem immediately : )

[access2rohit]

Thank you @mseth10 for looking into this !!
I think Option 2 helps us unblock CI thus unblocking all open source contributors immediately and gives us more time to take the corrective measure to refactor files to reduce memory utilization by gcc.

Option1 may keep us blocked for long and has the uncertainty of how long will it take.

IMHO Option 2 is the way forward :)
@leezu wdyt ?

[Zha0q1]

+1 for option 2

[kshitij12345]

A tangential question:

Should we raise a bug report with gcc community as it seems that gcc8 is using more resources to handle the same files than gcc7 and it is a regression between those version.

[leezu]

How do you arrive at the time estimates? Adding a CC=gcc-7 statement in runtime-functions.sh shouldn't take you 2 days. It should take less than 30 minutes... I'm also a bit worried about a time estimate of 40 days for splitting .cc files into multiple files.

[samskalicky]

How do you arrive at the time estimates? Adding a CC=gcc-7 statement in runtime-functions.sh shouldn't take you 2 days. It should take less than 30 minutes... I'm also a bit worried about a time estimate of 40 days for splitting .cc files into multiple files.

Effort will certainly vary depending on who does the work, their familiarity, and how long it takes them to ramp up. If we get people familiar with the code it could be shorter. If we get people unfamiliar, 5/day seems reasonable. If its someone new to MXNet (and possibly not an expert in C++) it could take longer.

Until we get actual people to volunteer to do the work, and ask them how long they will need to do the work, we cant really give hard estimates. So these will have to do for now.