Dataloader is slow with iterdatapipes and shuffle that has large in-memory fields (because traverse_dps is slow)

[olegsinavski]

🐛 Describe the bug

Hello, I found that a standard DataLoader takes unreasonably long to construct itself and to load the first batch if there is a filed in a dataset that takes long to pickle (e.g. an in-memory dataset with panda frame and strings).

This happens only if shuffle is true and the datapipe is an IterDataPipe. The dataloader calls apply_shuffle_settings which in turn calls traverse_dps, then _list_connected_datapipes which eventually pickles all object fields in a dataset. I was not able to comprehend why would one need to pickle datafields to build a datapipe graph.

Here is a reproduction:

import torch
import time
import pandas as pd
import numpy as np
from torchdata.datapipes.map import MapDataPipe

class TabularDataset(MapDataPipe):
    """ Make MapDataPipe so that shuffle is fast"""
    def __init__(self, data):
        self._data = data

    def __getitem__(self, idx: int):
        return  self._data['complex_objects'].iloc[idx]

    def __len__(self):
        return len(self._data)

# Large panda frames with strings take a long time to pickle
n_rows = 1000000
data = pd.DataFrame({
    'complex_objects': [
        np.array(["1", "2", "3", "4"]) 
        if r % 2 == 0 else np.array(["4", "3", "2", "1"])
        for r in range(n_rows)
    ]
})

dataset = TabularDataset(data).shuffle()
start_time = time.time()
dataloader = torch.utils.data.DataLoader(
    dataset, shuffle=True, batch_size=2, num_workers=0, collate_fn=lambda x: x
)
print(f"Dataloader creation time: {time.time() - start_time:.2f}s")
for b in dataloader:
    assert len(b) == 2
    print(b)
    print(f"Time to get the first batch: {time.time() - start_time:.2f}s")
    break

This code prints out:

Dataloader creation time: 7.44s
[array(['4', '3', '2', '1'], dtype='<U1'), array(['4', '3', '2', '1'], dtype='<U1')]
Time to get the first batch: 22.85s

In my case, I have even slower datapipe that takes 5 minutes to pickle.

A workaround

One can make a datapipe that doesn't take long to pickle (e.g. with lambdas):

class TabularDatasetIndirect(MapDataPipe):
    """ Indirect access to data"""
    def __init__(self, data):
        self._data_f = lambda: data

    def __getitem__(self, idx: int):
        return  self._data_f()['complex_objects'].iloc[idx]

    def __len__(self):
        return len(self._data_f())

In that case, the printout is

Dataloader creation time: 0.16s
[array(['1', '2', '3', '4'], dtype='<U1'), array(['4', '3', '2', '1'], dtype='<U1')]
Time to get the first batch: 0.75s

which is a 30x speedup in this case.

Versions

PyTorch version: 2.1.2
Is debug build: False
CUDA used to build PyTorch: 11.8
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Clang version: 13.0.1-++20220120110924+75e33f71c2da-1~exp1~20220120231001.58
CMake version: version 3.16.3
Libc version: glibc-2.31

Python version: 3.9.5 (default, Nov 23 2021, 15:27:38)  [GCC 9.3.0] (64-bit runtime)
Python platform: Linux-5.15.0-1038-azure-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: 11.8.89
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA A100-SXM4-80GB
GPU 1: NVIDIA A100-SXM4-80GB
GPU 2: NVIDIA A100-SXM4-80GB
GPU 3: NVIDIA A100-SXM4-80GB
GPU 4: NVIDIA A100-SXM4-80GB
GPU 5: NVIDIA A100-SXM4-80GB
GPU 6: NVIDIA A100-SXM4-80GB
GPU 7: NVIDIA A100-SXM4-80GB

Nvidia driver version: 535.129.03
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.5
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Byte Order:                      Little Endian
Address sizes:                   48 bits physical, 48 bits virtual
CPU(s):                          96
On-line CPU(s) list:             0-95
Thread(s) per core:              1
Core(s) per socket:              48
Socket(s):                       2
NUMA node(s):                    4
Vendor ID:                       AuthenticAMD
CPU family:                      23
Model:                           49
Model name:                      AMD EPYC 7V12 64-Core Processor
Stepping:                        0
CPU MHz:                         2445.440
BogoMIPS:                        4890.88
Hypervisor vendor:               Microsoft
Virtualization type:             full
L1d cache:                       3 MiB
L1i cache:                       3 MiB
L2 cache:                        48 MiB
L3 cache:                        384 MiB
NUMA node0 CPU(s):               0-23
NUMA node1 CPU(s):               24-47
NUMA node2 CPU(s):               48-71
NUMA node3 CPU(s):               72-95
Vulnerability Itlb multihit:     Not affected
Vulnerability L1tf:              Not affected
Vulnerability Mds:               Not affected
Vulnerability Meltdown:          Not affected
Vulnerability Mmio stale data:   Not affected
Vulnerability Retbleed:          Mitigation; untrained return thunk; SMT disabled
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:        Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:        Mitigation; Retpolines, STIBP disabled, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Not affected
Flags:                           fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl tsc_reliable nonstop_tsc cpuid extd_apicid aperfmperf pni pclmulqdq ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw topoext perfctr_core ssbd vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves clzero xsaveerptr rdpru arat umip rdpid

Versions of relevant libraries:
[pip3] efficientnet-pytorch==0.7.1
[pip3] numpy==1.25.2
[pip3] onnx==1.15.0
[pip3] open-clip-torch==2.19.0
[pip3] pytorch-lightning==2.1.2
[pip3] torch==2.1.2
[pip3] torch-fidelity==0.3.0
[pip3] torchdata==0.7.1
[pip3] torchmetrics==1.3.0.post0
[pip3] torchvision==0.16.1
[pip3] triton==2.2.0
[conda] Could not collect

cc @SsnL @VitalyFedyunin @ejguan @dzhulgakov

[ejguan]

We rely on pickle to help us traverse the graph because we don't know which datafield in your custom DataPipe contains a dependent DataPipe. It definitely can be optimized by skipping the datafield that is not any of the following types: DataPipe, list, tuple, dict, set..

[olegsinavski]

Thank you for the response @ejguan. Why would serializing something allows you to figure out the structure? I don't think it's designed for this purpose. Would just a regular python recursive field introspection work?

[ejguan]

In theory, we injected logics into pickle so that only traversal happens (serialization is not really happening). We did have some plan to optimize the performance toward traversal. But, unfortunately, we are not actively supporting this project anymore.