Dask improvements #208

rosepearson · 2023-10-11T03:59:27Z

This is an issue for an upcoming NeSI consultancy with @jennan. The focus will be on improving the performance and stability of GeoFabrics for larger scale problems.

Focus on better making use of Dask throughout the GeoFabrics stages. Two identified areas are:

RasterArray.interpolate_na for 2D interpolation to remove NaN values
RasterArray.clip in series for identifying the lowest value along a waterways

RasterArray.interpolate_na
Profiling has shown that pinch points are quite different for larger scale problems than smaller problems. Take the two profiles below.

is a 6min problem with all geofabric stages. (small_2m_res.html)
is a 4hr problem with all geograbric stages. The only difference with 1 is it is 1m instead of 2m resolution. (small_1m_res.html)

RasterArray.clip
Another area of focus (although it hasn't showed up as an issue in the 1m profiling it is vivible in the 2m) is

Could make use of either pandas or dask-geopandas. A weak attempt that I didn't get off the ground can be seen as a comment in processor.py

Also worth noting that we may be able to do this more directly using a rolling.min call to the xarray with an appropriate size window.

rosepearson · 2023-10-13T01:38:43Z

Have also noticed some unexpected failures with Dask TimeOut/Heartbeat errors. In some instances this is actually after the final chunk appears to have been written to file.

rosepearson · 2023-10-13T01:39:54Z

job.err.txt
Attached Cylc job.err file illustrating the Dask TimeOut/Heartbeat errors.

jennan · 2023-10-13T01:57:07Z

@rosepearson is it making the whole code crash? Dask can be a bit verbose with errors during the tearing down procedure of the Dask cluster... although this doesn't cause any real issue (apart from logs reporting errors) in my experience.

rosepearson · 2023-10-13T03:18:53Z

@jennan it is making the whole Cylc task crash. I have another one below where we hit the time limit - but again the file had been completely written out. It could be a coincident that the file just finished writing before the time limit was reached and it didn't have time to execute the print "Job Succeeded".. but it seems to me like there might be something slightly fishy. Just attaching as a reference for now - not expecting any action on this ticket right now :)

rosepearson · 2023-10-13T03:20:42Z

job.err.txt
Attached another Cylc job.err file. Again lots of verbose Task errors. The task ultimately failed after hitting the time limit - it had written out a complete file prior to hitting the time limit, however.

rosepearson · 2023-10-15T22:53:41Z

Noting another repeated error that occurs for a particular roughness run. It repeatedly fails when launched on Maui through Cylc, but has completed producing the expected output. e.e. appears to be a false failure. I have checked running on the NIWA maui nodes without SLURM accessed through https://jupyter.maui.niwa.co.nz/ and it runs fine there. This might be a good first place to look into regarding seemingly random errors after the job has run successfully.

job.err.txt

tile 2728. roughness stage. takes ~15min.

rosepearson · 2023-10-27T02:09:36Z

One more note about the failures I'm experiencing. They are often related to Sending large graph of size XX.XX MiB.

One is copied below. I am wondering if we should restructure how we deal with upsampling the coarse DEM to do that directly instead of breaking into chunks and doing each chunk individually. This would mean many fewer explicit dask delayed calls and leave it up to Dask exactly how it chooses to manage the compute load. This would mean the upsampling is all done with linear interpolation - but that seems sensible anyway.

/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/client.py:3160: UserWarning: Sending large graph of size 12.85 MiB.
This may cause some slowdown.
Consider scattering data ahead of time and using futures.
  warnings.warn(
2023-10-14 11:35:39,162 - distributed.protocol.core - CRITICAL - Failed to Serialize
Traceback (most recent call last):
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/protocol/core.py", line 109, in dumps
    frames[0] = msgpack.dumps(msg, default=_encode_default, use_bin_type=True)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/msgpack/__init__.py", line 38, in packb
    return Packer(**kwargs).pack(o)
           ^^^^^^^^^^^^^^^^^^^^^^^^
  File "msgpack/_packer.pyx", line 294, in msgpack._cmsgpack.Packer.pack
  File "msgpack/_packer.pyx", line 300, in msgpack._cmsgpack.Packer.pack
  File "msgpack/_packer.pyx", line 297, in msgpack._cmsgpack.Packer.pack
  File "msgpack/_packer.pyx", line 264, in msgpack._cmsgpack.Packer._pack
  File "msgpack/_packer.pyx", line 231, in msgpack._cmsgpack.Packer._pack
  File "msgpack/_packer.pyx", line 264, in msgpack._cmsgpack.Packer._pack
  File "msgpack/_packer.pyx", line 272, in msgpack._cmsgpack.Packer._pack
ValueError: memoryview is too large
2023-10-14 11:35:39,163 - distributed.comm.utils - ERROR - memoryview is too large

rosepearson · 2023-11-01T02:31:43Z

SUMMARY

Looking back through my notes/comments:

Tile 2728. roughness stage. takes taking ~15min then failing with the file completed.
Tile 2627 roughness stage, fialed with a memory error. See high transfer load in comment below
An ValueError: memoryview is too large error occurs for the waterway stage of tiles 3128, 3130, 3131, 3532
An ValueError: 3615135876 exceeds max_bin_len(2147483647) error occurs in the waterway stage for 3431

I've also been thinking about how I break up the course DEM stage into a bunch of explicit chunks and wonder if that would be best to leave up to Dask.

Previously load the same coarse .tif tile many times clipping it to a very small size each time and interpolating points pulled from the XArray
I could/should? instead load with chunking set to true and use the XArray interp function to upsample with linear interpolation. I then have a raster mask I would use to decide where to replace the DEM with these new values using xarray.where()
- See https://github.com/rosepearson/GeoFabrics/tree/208-dask-improvements for an attempt at making this change

Old relevant issues

Dask memory view error - fixed - ValueError: memoryview is too large; when reading from a large memory mapped file in Dask. dask/dask#7552
Dask large graph - distributed fails with large graph/dataset dask/distributed#4678

rosepearson · 2023-11-03T00:49:45Z

Tile 2627 roughness

rosepearson · 2023-11-06T03:43:41Z

Update after Consultancy 1

Most recent commit associated with this comment ae30061
Have made some of the agreed changes:

Save out LiADR only temp netCDF
Optionally save the no_data_mask to the netCDF
Load in the coarse DEM without chunking

I've been looking into xarray.interp with chunking. I've come across the following cases:

closed - interpolate a dask-based xarray between chunk of data- Feature request: Implement interp for interpolating between chunks of data (dask) pydata/xarray#4078
closed - Implement interp for interpolating between chunks of data - Implement interp for interpolating between chunks of data (dask) pydata/xarray#4155
open - dask implementation of np.interp - dask implementation of np.interp dask/dask#6474

I've also tried various ways to force/encourage dask in the interp call:

coarse_dem = coarse_dem.interp(x=dask.array.from_array(self._dem.x, 1000), y=dask.array.from_array(self._dem.y), method="linear")
- Error - in _validate_interp_indexers raise TypeError(type(v))
coarse_dem.chunk(1000).interp(x=dask.array.self._dem.x, y=self._dem.y, method="linear")
- No error, but result has only one chunk. Get memory errors... but runs for a bit before failure
I tried with both the normal chunk size and also one that was 8x smaller given the coarse resolution is 8x bigger.
It started saving out a file, but did not succeed in completely writing the file.

Questions

Is xarray.interp implemented with dask support - see dashboard screen captures it seems to run across cores
Should we chunk the coarse_dem after all as chunking the x/y dims on their own don't seem to work...

rosepearson · 2023-11-08T22:37:02Z

@jennan thanks for your notebook. I've implemented it with a few minor tweaks to deal with the std map layout (yx with y decreasing). Also a seemingly odd issue where the chunking needs dask.array.map_blocks needs the second array to have a chunk of its length or smaller - odd at the first array is autmatically given a chunk of its size if the specified chunk size is greater.

No action required - just an update

Test profiling for tests/test_dem_generation_westport_4/test_dem_generation_westport_4.py
top chunking to 10 - gives 4x5 chunks (1m54s)

where the
bottom xhunking to 300 - gives 1x1 chunks (1m15s)

Testing on (previously failing larger example.

It ran successfully produced the netCDF file which is new. It did fail with a TimeOut error shortly afterwards.

2023-11-10 02:59:37,360 - distributed.nanny - WARNING - Worker process still alive after 3.199999694824219 seconds, killing
2023-11-10 02:59:38,137 - tornado.application - ERROR - Exception in callback functools.partial(<bound method IOLoop._discard_future_result of <tornado.platform.asyncio.AsyncIOMainLoop object at 0x2aad57c94d90>>, <Task finished name='Task-180509187' coro=<SpecCluster._correct_state_internal() done, defined at /nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/deploy/spec.py:346> exception=TimeoutError()>)
Traceback (most recent call last):
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/utils.py", line 1922, in wait_for
    return await fut
           ^^^^^^^^^
asyncio.exceptions.CancelledError

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/tornado/ioloop.py", line 738, in _run_callback
    ret = callback()
          ^^^^^^^^^^
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/tornado/ioloop.py", line 762, in _discard_future_result
    future.result()
TimeoutError
Traceback (most recent call last):
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/utils.py", line 1922, in wait_for
    return await fut
           ^^^^^^^^^
asyncio.exceptions.CancelledError

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/scale_wlg_persistent/filesets/project/niwa03440/geofabrics/GeoFabrics/src/geofabrics/__main__.py", line 53, in <module>
    cli_run_from_file()
  File "/scale_wlg_persistent/filesets/project/niwa03440/geofabrics/GeoFabrics/src/geofabrics/__main__.py", line 47, in cli_run_from_file
    runner.from_instructions_file(instructions_path=args.instructions)
  File "/scale_wlg_persistent/filesets/project/niwa03440/geofabrics/GeoFabrics/src/geofabrics/runner.py", line 211, in from_instructions_file
    from_instructions_dict(instructions=instructions)
  File "/scale_wlg_persistent/filesets/project/niwa03440/geofabrics/GeoFabrics/src/geofabrics/runner.py", line 153, in from_instructions_dict
    run_processor_class(
  File "/scale_wlg_persistent/filesets/project/niwa03440/geofabrics/GeoFabrics/src/geofabrics/runner.py", line 48, in run_processor_class
    runner.run()
  File "/scale_wlg_persistent/filesets/project/niwa03440/geofabrics/GeoFabrics/src/geofabrics/processor.py", line 2999, in run
    dem = self.create_dem(waterways=waterways)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/scale_wlg_persistent/filesets/project/niwa03440/geofabrics/GeoFabrics/src/geofabrics/processor.py", line 2861, in create_dem
    runner.run()
  File "/scale_wlg_persistent/filesets/project/niwa03440/geofabrics/GeoFabrics/src/geofabrics/processor.py", line 910, in run
    with cluster, distributed.Client(cluster) as client:
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/deploy/cluster.py", line 540, in __exit__
    aw = self.close()
         ^^^^^^^^^^^^
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/deploy/spec.py", line 293, in close
    aw = super().close(timeout)
         ^^^^^^^^^^^^^^^^^^^^^^
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/deploy/cluster.py", line 226, in close
    return self.sync(self._close, callback_timeout=timeout)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/utils.py", line 359, in sync
    return sync(
           ^^^^^
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/utils.py", line 426, in sync
    raise exc.with_traceback(tb)
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/utils.py", line 399, in f
    result = yield future
             ^^^^^^^^^^^^
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/tornado/gen.py", line 767, in run
    value = future.result()
            ^^^^^^^^^^^^^^^
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/deploy/spec.py", line 446, in _close
    await self._correct_state()
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/deploy/spec.py", line 359, in _correct_state_internal
    await asyncio.gather(*tasks)
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/nanny.py", line 595, in close
    await self.kill(timeout=timeout, reason=reason)
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/nanny.py", line 380, in kill
    await self.process.kill(reason=reason, timeout=0.8 * (deadline - time()))
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/nanny.py", line 843, in kill
    await process.join(max(0, deadline - time()))
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/process.py", line 330, in join
    await wait_for(asyncio.shield(self._exit_future), timeout)
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/site-packages/distributed/utils.py", line 1921, in wait_for
    async with asyncio.timeout(timeout):
  File "/nesi/project/niwa03440/conda/envs/geofabrics/lib/python3.11/asyncio/timeouts.py", line 111, in __aexit__
    raise TimeoutError from exc_val
TimeoutError

dashboard during LiDAR write

dashboard at end of execution

rosepearson · 2023-11-13T03:49:35Z

Update from consultancy 2

Pulled changes from first consultancy into main - See Issue 219
Running the Cylc suite over ~40 tiles (of 425)
Record - Success/Failure and run time for successful execution compared to last time.
timing_comparison.xlsx

Identified test cases

Time-limit/large graph size - waterways 2827, 3432
i. I'll rerun this without a timelimit and see how long it takes or if we get a different error
Failure to close Dask cluster after completing the file write - waterways 2113, 2423, 2728, 3232, lidar 2828
i. Example error log at: /home/pearsonra/cylc-run/cylc-geofabrics/run16/log/job/1/waterways_tile_id3232/01/job.err

Separate bugfix to fix regression errors

Issue 220
Fixes lidar 2112
Fixes tiles 1514, 1614, 1615, 2827, 3027, 3331

Next consultancy

Would be great to discuss how to address the failure at client closure errors
45min testcase ending in client close error: /nesi/project/niwa03440/Cylc-Workflow-Outputs/geofabrics-tiles/tiles/2828/instruction_lidar.json
@jennan let me know if there is anything related to this I should prepare prior to Monday

Error logs

File write completed, Dask timeout - /home/personra/cylc-run/cylc-geofabrics/run16/log/job/1/lidar_tile_id2828/01/job.err 45min run to crash
File write completed, Dask timeout - /home/personra/cylc-run/cylc-geofabrics/run16/log/job/1/waterways_tile_id3532/01/job.err
File write completed, Dask timeout - /home/personra/cylc-run/cylc-geofabrics/run16/log/job/1/waterways_tile_id3232/01/job.err
File write completed, Dask timeout - /home/personra/cylc-run/cylc-geofabrics/run16/log/job/1/waterways_tile_id3138/01/job.err
File write completed, Dask timeout - /home/personra/cylc-run/cylc-geofabrics/run16/log/job/1/waterways_tile_id2728/01/job.err
File write completed, Dask timeout - /home/personra/cylc-run/cylc-geofabrics/run16/log/job/1/waterways_tile_id2423/01/job.err
File write completed, Dask timeout - /home/personra/cylc-run/cylc-geofabrics/run16/log/job/1/waterways_tile_id2113/01/job.err

rosepearson · 2023-11-19T20:59:49Z

Consultancy 3

Tasks for Rose

Collect logs into a folder under niwa03440
Check if increasing the retry count helps - export DASK_DISTRIBUTED__COMM__RETRY__COUNT=3
Find the Dask folder is created - dask-worker-space - found in tmp\dask-worker-space
Update processor.py to nest context for clusters and then client. Tried - still an error.
Update processor.py cluster variable with dashboard_address: None. Tried - still an error.
Update processor.py cluster variable with "local_directory": "path_to_folder_with_results. This produces an output in the folder - a global.lock & purge.lock file created and not deleted prior to crash.

rosepearson linked a pull request Nov 2, 2023 that will close this issue

Dask improvements #217

Merged

10 tasks

rosepearson mentioned this issue Nov 12, 2023

Dask Improvements part 1 #219

Closed

rosepearson removed a link to a pull request Nov 12, 2023

Dask improvements #217

Merged

10 tasks

rosepearson linked a pull request Nov 30, 2023 that will close this issue

bug-fixes identified in cylc suite #222

Merged

10 tasks

rosepearson removed a link to a pull request Nov 30, 2023

bug-fixes identified in cylc suite #222

Merged

10 tasks

rosepearson added this to the 1.1.6 milestone Jan 4, 2024

rosepearson linked a pull request Jan 4, 2024 that will close this issue

Refactor rawdem #224

Merged

10 tasks

rosepearson closed this as completed in #224 Jan 4, 2024

rosepearson mentioned this issue Jan 5, 2024

Roughness stage Dask optimisations #229

Closed

5 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Dask improvements #208

Dask improvements #208

rosepearson commented Oct 11, 2023

rosepearson commented Oct 13, 2023

rosepearson commented Oct 13, 2023

jennan commented Oct 13, 2023

rosepearson commented Oct 13, 2023

rosepearson commented Oct 13, 2023

rosepearson commented Oct 15, 2023

rosepearson commented Oct 27, 2023 •

edited

Loading

rosepearson commented Nov 1, 2023 •

edited

Loading

rosepearson commented Nov 3, 2023

rosepearson commented Nov 6, 2023 •

edited

Loading

rosepearson commented Nov 8, 2023 •

edited

Loading

rosepearson commented Nov 13, 2023 •

edited

Loading

rosepearson commented Nov 19, 2023 •

edited

Loading

Dask improvements #208

Dask improvements #208

Comments

rosepearson commented Oct 11, 2023

rosepearson commented Oct 13, 2023

rosepearson commented Oct 13, 2023

jennan commented Oct 13, 2023

rosepearson commented Oct 13, 2023

rosepearson commented Oct 13, 2023

rosepearson commented Oct 15, 2023

rosepearson commented Oct 27, 2023 • edited Loading

rosepearson commented Nov 1, 2023 • edited Loading

SUMMARY

Old relevant issues

rosepearson commented Nov 3, 2023

rosepearson commented Nov 6, 2023 • edited Loading

Update after Consultancy 1

Questions

rosepearson commented Nov 8, 2023 • edited Loading

Testing on (previously failing larger example.

rosepearson commented Nov 13, 2023 • edited Loading

Update from consultancy 2

Identified test cases

Separate bugfix to fix regression errors

Next consultancy

Error logs

rosepearson commented Nov 19, 2023 • edited Loading

Consultancy 3

Tasks for Rose

rosepearson commented Oct 27, 2023 •

edited

Loading

rosepearson commented Nov 1, 2023 •

edited

Loading

rosepearson commented Nov 6, 2023 •

edited

Loading

rosepearson commented Nov 8, 2023 •

edited

Loading

rosepearson commented Nov 13, 2023 •

edited

Loading

rosepearson commented Nov 19, 2023 •

edited

Loading