Unknown error when using this in PyTorch DDP

[bchuh]

I‘m doing Pytorch distributed data parallel training, and I used a generator to traverse the EArray data in my hdf5 dataset. As soon as the program started, I got an "RuntimeError: Unknown error type: 132 when handling execution of <_FuncPtr object at 0x7fb033f58280> with args (b'/nXmxo38xgBw=', 194, 384)". The code where multitables is involved is shown below.

class SignalTrainSetSpeedy(Dataset):

    def __init__(self, augment='None'):
        self.aug = augment
        self.input_path = "/mnt/sdb/zzl/train_data_and_label.hdf5"
        self.stream = None
        self.data = None

    def _yield_data(self):
        for row in self.gen:
            yield row

    def __getitem__(self, index):
        if self.stream is None:
            self.stream = multitables.Streamer(filename=self.input_path)
            self.gen = self.stream.get_generator(path='/data')
            self.data = self._yield_data()
        input_output = next(self.data)
        input = input_output[:16000]
        target = input_output[-61:]
        input, target = torch.Tensor(input), torch.Tensor(target)
        return input, target

    def __len__(self):
        return 2040151

The error message:

Traceback (most recent call last):
  File "train.py", line 244, in <module>
    next_batch = dataset_iter.next()
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 530, in __next__
Traceback (most recent call last):
  File "train.py", line 244, in <module>
    data = self._next_data()
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 570, in _next_data
    next_batch = dataset_iter.next()
      File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 530, in __next__
data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/mnt/sdf/zzl/FACEGOOD-Audio2Face/code/torch_train/signal_dataset.py", line 114, in __getitem__
    data = self._next_data()
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 570, in _next_data
    input_output = next(self.data)
  File "/mnt/sdf/zzl/FACEGOOD-Audio2Face/code/torch_train/signal_dataset.py", line 103, in _yield_data
    for row in self.gen:
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/streamer.py", line 247, in get_generator
    data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in fetch
    q = self.get_queue(path=path, n_procs=n_procs, read_ahead=read_ahead, cyclic=cyclic, block_size=block_size, ordered=ordered, field=field, remainder=remainder)
      File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/streamer.py", line 195, in get_queue
data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/mnt/sdf/zzl/FACEGOOD-Audio2Face/code/torch_train/signal_dataset.py", line 114, in __getitem__
    stage_pool = stage.StagePool(dataset, block_size, read_ahead, timeout=0.1)
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/stage.py", line 104, in __init__
        self._stage_pool.append(StagePool.StagePoolWrapper(dataset.create_stage(stage_size), self))input_output = next(self.data)

  File "/mnt/sdf/zzl/FACEGOOD-Audio2Face/code/torch_train/signal_dataset.py", line 103, in _yield_data
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/dataset.py", line 64, in create_stage
    return stage.Stage(numpy_utils._calc_nbytes(self.dtype, self._fill_shape(shape)))
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/stage.py", line 29, in __init__
    for row in self.gen:
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/streamer.py", line 247, in get_generator
    self._shm_buf = shared_mem.SharedBuffer(map_id=None, size_nbytes=self.size_nbytes)
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/shared_mem.py", line 124, in __init__
    q = self.get_queue(path=path, n_procs=n_procs, read_ahead=read_ahead, cyclic=cyclic, block_size=block_size, ordered=ordered, field=field, remainder=remainder)
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/streamer.py", line 195, in get_queue
    self._fd = _shm_open(map_id, os.O_CREAT | os.O_EXCL | os.O_RDWR, mode=0o600)
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/shared_mem.py", line 64, in _shm_open
        stage_pool = stage.StagePool(dataset, block_size, read_ahead, timeout=0.1)return _posixshmlib.shm_open(name, access, mode)

  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/stage.py", line 104, in __init__
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/shared_mem.py", line 46, in _handle_errno
    raise RuntimeError("Unknown error type: {} when handling execution of {} with args {}".format(erno, func, args))
    self._stage_pool.append(StagePool.StagePoolWrapper(dataset.create_stage(stage_size), self))
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/dataset.py", line 64, in create_stage
RuntimeError: Unknown error type: 132 when handling execution of <_FuncPtr object at 0x7f00263a2280> with args (b'/nXmxo38xgBw=', 194, 384)
    return stage.Stage(numpy_utils._calc_nbytes(self.dtype, self._fill_shape(shape)))
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/stage.py", line 29, in __init__
    self._shm_buf = shared_mem.SharedBuffer(map_id=None, size_nbytes=self.size_nbytes)
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/shared_mem.py", line 124, in __init__
    self._fd = _shm_open(map_id, os.O_CREAT | os.O_EXCL | os.O_RDWR, mode=0o600)
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/shared_mem.py", line 64, in _shm_open
    return _posixshmlib.shm_open(name, access, mode)
  File "/home/zhuzengliang/anaconda3/envs/pytorch1/lib/python3.8/site-packages/multitables/shared_mem.py", line 46, in _handle_errno
    raise RuntimeError("Unknown error type: {} when handling execution of {} with args {}".format(erno, func, args))
RuntimeError: Unknown error type: 132 when handling execution of <_FuncPtr object at 0x7f832a6a3280> with args (b'/nXmxo38xgBw=', 194, 384)

[ghcollin]

Unfortunately this will be hard to diagnose, as the error is coming from the OS and python does not seem to know what the error code corresponds to.

I can only guess that because this is happening when the shared memory is being allocated, and because in some POSIX error mappings, 132 corresponds to EOVERFLOW, that the problem might be that multitables is trying to allocate more shared memory than is allowed on your system.

Multitables defaults to allocating the chunk size of the array, maybe your chunk size is too large? Try setting block_size=1 when calling get_generator to reduce the amount allocated, or increasing the amount of allowed shared memory in your OS settings.

[bchuh]

Adding the "block_size=1" arg. does not solve it. I also tried switching to non-parallel training, but the error remains the same.
I guess I'm out of luck here, but thanks for replying anyway!