Skip to content

Instantly share code, notes, and snippets.

@jdye64

jdye64/segfault-investigation.md

Last active October 29, 2021 01:04
Show Gist options
  • Save jdye64/dfb7b3a9a7bc186777a64118d042b90d to your computer and use it in GitHub Desktop.
Save jdye64/dfb7b3a9a7bc186777a64118d042b90d to your computer and use it in GitHub Desktop.
Download ZIP
Raw
segfault-investigation.md

Dask-SQL SegFault notes and observations

Notes

  • Occurs regardless of LocalCUDACluster transport specified. Ex: UCX, TCP, etc
  • Only occurs when ucx-py is installed in the Anaconda environment AND LocalCUDACluster is used instead of standard Distributed.Client
  • Any environment without UCX and issues cannot be reproduced

Results

I have provided 2 test cases. One with ucx and one without. The tests are as close as possible (some imports had to be removed) to demonstrate the failures.

dask-sql-no-ucx

Conda

conda create --name dask-sql-no-ucx
conda activate dask-sql-no-ucx
conda install -c rapidsai-nightly -c nvidia -c numba -c conda-forge cudf dask-cudf dask-cuda  python=3.7 cudatoolkit=11.2 openjdk maven
python ./setup.py install // Assuming you are in the dask-sql repo directory

Minimal Reproducer

from dask.distributed import Client
import dask_cudf as dd

import cudf
from dask_sql import Context

if __name__ == "__main__":
    client = Client()
    client

    c = Context()
    df = cudf.DataFrame({'id': [0, 1]})
    c.create_table('test', df)
    print(c.sql("select count(*) from test").compute())

Output

   COUNT(*)
0         2

dask-sql

Conda

conda create --name dask-sql 
conda activate dask-sql
conda install -c rapidsai-nightly -c nvidia -c numba -c conda-forge cudf dask-cudf dask-cuda python=3.7 cudatoolkit=11.2 openjdk maven ucx-py ucx-proc=*=gpu
python ./setup.py install // Assuming you are in the dask-sql repo directory

Minimal Reproducer

from dask.distributed import Client
import dask_cudf as dd

import cudf
from dask_sql import Context
from dask_cuda import LocalCUDACluster


if __name__ == "__main__":
    cluster = LocalCUDACluster(protocol="ucx", enable_tcp_over_ucx=True, enable_nvlink=True, jit_unspill=False, rmm_pool_size="29GB")
    client = Client(cluster)
    client

    c = Context()
    df = cudf.DataFrame({'id': [0, 1]})
    c.create_table('test', df)
    print(c.sql("select count(*) from test").compute())

Output

Note that the JVM process is not causing the actual SIGSEGV but rather is providing a core dump as the result of one. The parent process seems to be the guilty party but then the JVM output shows up in the output once the parent process is killed by the kernel and the JVM prints output like that when either its process OR parent process receives a SIGSEGV

distributed.preloading - INFO - Import preload module: dask_cuda.initialize
distributed.preloading - INFO - Import preload module: dask_cuda.initialize
distributed.preloading - INFO - Import preload module: dask_cuda.initialize
distributed.preloading - INFO - Import preload module: dask_cuda.initialize
distributed.preloading - INFO - Import preload module: dask_cuda.initialize
distributed.preloading - INFO - Import preload module: dask_cuda.initialize
distributed.preloading - INFO - Import preload module: dask_cuda.initialize
distributed.preloading - INFO - Import preload module: dask_cuda.initialize
[rl-dgx-r13-u24-rapids-dgx118:38784:0:38784] Caught signal 11 (Segmentation fault: invalid permissions for mapped object at address 0x7f7e2d26b008)
==== backtrace (tid:  38784) ====
 0  /home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/ucp/_libs/../../../../libucs.so.0(ucs_handle_error+0x115) [0x7f7d7e2cf4e5]
 1  /home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/ucp/_libs/../../../../libucs.so.0(+0x2a881) [0x7f7d7e2cf881]
 2  /home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/ucp/_libs/../../../../libucs.so.0(+0x2aa52) [0x7f7d7e2cfa52]
 3  [0x7f7acfb4144d]
=================================
#
# A fatal error has been detected by the Java Runtime Environment:
#
#  SIGSEGV (0xb) at pc=0x00007f7acfb4144d (sent by kill), pid=38784, tid=38784
#
# JRE version: OpenJDK Runtime Environment (11.0.9.1) (build 11.0.9.1-internal+0-adhoc..src)
# Java VM: OpenJDK 64-Bit Server VM (11.0.9.1-internal+0-adhoc..src, mixed mode, tiered, compressed oops, g1 gc, linux-amd64)
# Problematic frame:
# J 1422 c1 java.util.WeakHashMap.put(Ljava/lang/Object;Ljava/lang/Object;)Ljava/lang/Object; [email protected] (162 bytes) @ 0x00007f7acfb4144d [0x00007f7acfb407a0+0x0000000000000cad]
#
# Core dump will be written. Default location: Core dumps may be processed with "/usr/share/apport/apport %p %s %c %d %P" (or dumping to /home/u00u9018xfl6yNnCjC357/development/dask-sql/core.38784)
#
# An error report file with more information is saved as:
# /home/u00u9018xfl6yNnCjC357/development/dask-sql/hs_err_pid38784.log
Compiled method (c1)    9329 1421       3       java.util.Collections$SetFromMap::add (22 bytes)
 total in heap  [0x00007f7acfb3eb90,0x00007f7acfb3f050] = 1216
 relocation     [0x00007f7acfb3ed08,0x00007f7acfb3ed48] = 64
 main code      [0x00007f7acfb3ed60,0x00007f7acfb3ef80] = 544
 stub code      [0x00007f7acfb3ef80,0x00007f7acfb3efc8] = 72
 metadata       [0x00007f7acfb3efc8,0x00007f7acfb3efd0] = 8
 scopes data    [0x00007f7acfb3efd0,0x00007f7acfb3efe8] = 24
 scopes pcs     [0x00007f7acfb3efe8,0x00007f7acfb3f038] = 80
 dependencies   [0x00007f7acfb3f038,0x00007f7acfb3f040] = 8
 nul chk table  [0x00007f7acfb3f040,0x00007f7acfb3f050] = 16
Compiled method (c1)    9330 1534       3       java.util.zip.ZipFile::getZipEntry (301 bytes)
 total in heap  [0x00007f7acfb86010,0x00007f7acfb88b78] = 11112
 relocation     [0x00007f7acfb86188,0x00007f7acfb86388] = 512
 main code      [0x00007f7acfb863a0,0x00007f7acfb88020] = 7296
 stub code      [0x00007f7acfb88020,0x00007f7acfb880e8] = 200
 metadata       [0x00007f7acfb880e8,0x00007f7acfb88158] = 112
 scopes data    [0x00007f7acfb88158,0x00007f7acfb88660] = 1288
 scopes pcs     [0x00007f7acfb88660,0x00007f7acfb88b00] = 1184
 dependencies   [0x00007f7acfb88b00,0x00007f7acfb88b08] = 8
 nul chk table  [0x00007f7acfb88b08,0x00007f7acfb88b78] = 112
Could not load hsdis-amd64.so; library not loadable; PrintAssembly is disabled
#
# If you would like to submit a bug report, please visit:
#   https://bugreport.java.com/bugreport/crash.jsp
#
distributed.worker - WARNING - Heartbeat to scheduler failed
Traceback (most recent call last):
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/distributed/comm/ucx.py", line 295, in read
    await self.ep.recv(msg)
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/ucp/core.py", line 725, in recv
    ret = await comm.tag_recv(self._ep, buffer, nbytes, tag, name=log)
ucp.exceptions.UCXCanceled: <[Recv #006] ep: 0x7f9b5c2500f0, tag: 0x8a71230ec78dfc21, nbytes: 16, type: <class 'numpy.ndarray'>>: 

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/distributed/worker.py", line 1197, in heartbeat
    for key in self.active_keys
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/distributed/utils_comm.py", line 390, in retry_operation
    operation=operation,
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/distributed/utils_comm.py", line 370, in retry
    return await coro()
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/distributed/core.py", line 863, in send_recv_from_rpc
    result = await send_recv(comm=comm, op=key, **kwargs)
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/distributed/core.py", line 640, in send_recv
    response = await comm.read(deserializers=deserializers)
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/distributed/comm/ucx.py", line 313, in read
    raise CommClosedError("Connection closed by writer")
distributed.comm.core.CommClosedError: Connection closed by writer
distributed.worker - WARNING - Heartbeat to scheduler failed
Traceback (most recent call last):
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/distributed/comm/ucx.py", line 295, in read
    await self.ep.recv(msg)
  File "/home/u00u9018xfl6yNnCjC357/miniconda3/envs/dask-sql/lib/python3.7/site-packages/ucp/core.py", line 725, in recv
    ret = await comm.tag_recv(self._ep, buffer, nbytes, tag, name=log)
ucp.exceptions.UCXCanceled: <[Recv #006] ep: 0x7f956041e0f0, tag: 0x5b546f00aaab3a53, nbytes: 16, type: <class 'numpy.ndarray'>>:
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment