Concatenate small input chunks before P2P rechunking (#8832)

Co-authored-by: Patrick Hoefler <61934744+phofl@users.noreply.github.com>

Author: hendrikmakait

distributed/shuffle/_rechunk.py

@@ -96,6 +96,7 @@
 
 from __future__ import annotations
 
+import math
 import mmap
 import os
 from collections import defaultdict
@@ -111,7 +112,7 @@
 )
 from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass
-from itertools import product
+from itertools import chain, product
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, NamedTuple, cast
 
@@ -124,6 +125,7 @@
 from dask.highlevelgraph import HighLevelGraph
 from dask.layers import Layer
 from dask.typing import Key
+from dask.utils import parse_bytes
 
 from distributed.core import PooledRPCCall
 from distributed.metrics import context_meter
@@ -220,7 +222,7 @@ def rechunk_p2p(
         return da.empty(x.shape, chunks=chunks, dtype=x.dtype)
     from dask.array.core import new_da_object
 
-    prechunked = _calculate_prechunking(x.chunks, chunks)
+    prechunked = _calculate_prechunking(x.chunks, chunks, x.dtype, block_size_limit)
     if prechunked != x.chunks:
         x = cast(
             "da.Array",
@@ -433,8 +435,140 @@ def _construct_graph(self) -> _T_LowLevelGraph:
 
 
 def _calculate_prechunking(
-    old_chunks: ChunkedAxes, new_chunks: ChunkedAxes
+    old_chunks: ChunkedAxes,
+    new_chunks: ChunkedAxes,
+    dtype: np.dtype,
+    block_size_limit: int | None,
+) -> ChunkedAxes:
+    """Calculate how to perform the pre-rechunking step
+
+    During the pre-rechunking step, we
+      1. Split input chunks along partial boundaries to make partials completely independent of one another
+      2. Merge small chunks within partials to reduce the number of transfer tasks and corresponding overhead
+    """
+    split_axes = _split_chunks_along_partial_boundaries(old_chunks, new_chunks)
+
+    # We can only determine how to concatenate chunks if we can calculate block sizes.
+    has_nans = (any(math.isnan(y) for y in x) for x in old_chunks)
+
+    if len(new_chunks) <= 1 or not all(new_chunks) or any(has_nans):
+        return tuple(tuple(chain(*axis)) for axis in split_axes)
+
+    if dtype is None or dtype.hasobject or dtype.itemsize == 0:
+        return tuple(tuple(chain(*axis)) for axis in split_axes)
+
+    # We made sure that there are no NaNs in split_axes above
+    return _concatenate_small_chunks(
+        split_axes, old_chunks, new_chunks, dtype, block_size_limit  # type: ignore[arg-type]
+    )
+
+
+def _concatenate_small_chunks(
+    split_axes: list[list[list[int]]],
+    old_chunks: ChunkedAxes,
+    new_chunks: ChunkedAxes,
+    dtype: np.dtype,
+    block_size_limit: int | None,
 ) -> ChunkedAxes:
+    """Concatenate small chunks within partials.
+
+    By concatenating chunks within partials, we reduce the number of P2P transfer tasks and their
+    corresponding overhead.
+
+    The algorithm used in this function is very similar to :func:`dask.array.rechunk.find_merge_rechunk`,
+    the main difference is that we have to make sure only to merge chunks within partials.
+    """
+    import numpy as np
+
+    block_size_limit = block_size_limit or dask.config.get("array.chunk-size")
+
+    if isinstance(block_size_limit, str):
+        block_size_limit = parse_bytes(block_size_limit)
+
+    # Make it a number of elements
+    block_size_limit //= dtype.itemsize
+
+    # We verified earlier that we do not have any NaNs
+    largest_old_block = _largest_block_size(old_chunks)  # type: ignore[arg-type]
+    largest_new_block = _largest_block_size(new_chunks)  # type: ignore[arg-type]
+    block_size_limit = max([block_size_limit, largest_old_block, largest_new_block])
+
+    old_largest_width = [max(chain(*axis)) for axis in split_axes]
+    new_largest_width = [max(c) for c in new_chunks]
+
+    # This represents how much each dimension increases (>1) or reduces (<1)
+    # the graph size during rechunking
+    graph_size_effect = {
+        dim: len(new_axis) / sum(map(len, split_axis))
+        for dim, (split_axis, new_axis) in enumerate(zip(split_axes, new_chunks))
+    }
+
+    ndim = len(old_chunks)
+
+    # This represents how much each dimension increases (>1) or reduces (<1) the
+    # largest block size during rechunking
+    block_size_effect = {
+        dim: new_largest_width[dim] / (old_largest_width[dim] or 1)
+        for dim in range(ndim)
+    }
+
+    # Our goal is to reduce the number of nodes in the rechunk graph
+    # by concatenating some adjacent chunks, so consider dimensions where we can
+    # reduce the # of chunks
+    candidates = [dim for dim in range(ndim) if graph_size_effect[dim] <= 1.0]
+
+    # Concatenating along each dimension reduces the graph size by a certain factor
+    # and increases memory largest block size by a certain factor.
+    # We want to optimize the graph size while staying below the given
+    # block_size_limit.  This is in effect a knapsack problem, except with
+    # multiplicative values and weights.  Just use a greedy algorithm
+    # by trying dimensions in decreasing value / weight order.
+    def key(k: int) -> float:
+        gse = graph_size_effect[k]
+        bse = block_size_effect[k]
+        if bse == 1:
+            bse = 1 + 1e-9
+        return (np.log(gse) / np.log(bse)) if bse > 0 else 0
+
+    sorted_candidates = sorted(candidates, key=key)
+
+    concatenated_axes: list[list[int]] = [[] for i in range(ndim)]
+
+    # Sim all the axes that are no candidates
+    for i in range(ndim):
+        if i in candidates:
+            continue
+        concatenated_axes[i] = list(chain(*split_axes[i]))
+
+    # We want to concatenate chunks
+    for axis_index in sorted_candidates:
+        concatenated_axis = concatenated_axes[axis_index]
+        multiplier = math.prod(
+            old_largest_width[:axis_index] + old_largest_width[axis_index + 1 :]
+        )
+        axis_limit = block_size_limit // multiplier
+
+        for partial in split_axes[axis_index]:
+            current = partial[0]
+            for chunk in partial[1:]:
+                if (current + chunk) > axis_limit:
+                    concatenated_axis.append(current)
+                    current = chunk
+                else:
+                    current += chunk
+            concatenated_axis.append(current)
+        old_largest_width[axis_index] = max(concatenated_axis)
+    return tuple(tuple(axis) for axis in concatenated_axes)
+
+
+def _split_chunks_along_partial_boundaries(
+    old_chunks: ChunkedAxes, new_chunks: ChunkedAxes
+) -> list[list[list[float]]]:
+    """Split the old chunks along the boundaries of partials, i.e., groups of new chunks that share the same inputs.
+
+    By splitting along the boundaries before rechunkin their input tasks become disjunct and each partial conceptually
+    operates on an independent sub-array.
+    """
     from dask.array.rechunk import old_to_new
 
     _old_to_new = old_to_new(old_chunks, new_chunks)
@@ -443,10 +577,13 @@ def _calculate_prechunking(
 
     split_axes = []
 
+    # Along each axis, we want to figure out how we have to split input chunks in order to make
+    # partials disjunct. We then group the resulting input chunks per partial before returning.
     for axis_index, slices in enumerate(partials):
         old_to_new_axis = _old_to_new[axis_index]
         old_axis = old_chunks[axis_index]
         split_axis = []
+        partial_chunks = []
         for slice_ in slices:
             first_new_chunk = slice_.start
             first_old_chunk, first_old_slice = old_to_new_axis[first_new_chunk][0]
@@ -465,22 +602,28 @@ def _calculate_prechunking(
                     chunk_size = last_old_slice.stop
                 if first_old_slice.start != 0:
                     chunk_size -= first_old_slice.start
-                split_axis.append(chunk_size)
-                continue
-
-            split_axis.append(first_chunk_size - first_old_slice.start)
-
-            split_axis.extend(old_axis[first_old_chunk + 1 : last_old_chunk])
-
-            if last_old_slice.stop is not None:
-                chunk_size = last_old_slice.stop
+                partial_chunks.append(chunk_size)
             else:
-                chunk_size = last_chunk_size
+                partial_chunks.append(first_chunk_size - first_old_slice.start)
 
-            split_axis.append(chunk_size)
+                partial_chunks.extend(old_axis[first_old_chunk + 1 : last_old_chunk])
 
+                if last_old_slice.stop is not None:
+                    chunk_size = last_old_slice.stop
+                else:
+                    chunk_size = last_chunk_size
+
+                partial_chunks.append(chunk_size)
+            split_axis.append(partial_chunks)
+            partial_chunks = []
+        if partial_chunks:
+            split_axis.append(partial_chunks)
         split_axes.append(split_axis)
-    return tuple(tuple(axis) for axis in split_axes)
+    return split_axes
+
+
+def _largest_block_size(chunks: tuple[tuple[int, ...], ...]) -> int:
+    return math.prod(map(max, chunks))
 
 
 def _split_partials(

distributed/shuffle/tests/test_rechunk.py

@@ -847,7 +847,8 @@ async def test_rechunk_avoid_needless_chunking(c, s, *ws):
     x = da.ones(16, chunks=2)
     y = x.rechunk(8, method="p2p")
     dsk = y.__dask_graph__()
-    assert len(dsk) <= 8 + 2
+    # 8 inputs, 2 concatenations of small inputs, 2 outputs
+    assert len(dsk) <= 8 + 2 + 2
 
 
 @pytest.mark.parametrize(
@@ -1337,7 +1338,7 @@ async def test_partial_rechunk_taskgroups(c, s):
         ),
         timeout=5,
     )
-    assert len(s.task_groups) < 6
+    assert len(s.task_groups) < 7
 
 
 @pytest.mark.parametrize(
@@ -1351,25 +1352,107 @@ async def test_partial_rechunk_taskgroups(c, s):
     ],
 )
 def test_calculate_prechunking_1d(old, new, expected):
-    actual = _calculate_prechunking(old, new)
+    actual = _calculate_prechunking(old, new, np.dtype, None)
     assert actual == expected
 
 
 @pytest.mark.parametrize(
     ["old", "new", "expected"],
     [
         [((2, 2), (3, 3)), ((2, 2), (3, 3)), ((2, 2), (3, 3))],
-        [((2, 2), (3, 3)), ((4,), (3, 3)), ((2, 2), (3, 3))],
+        [((2, 2), (3, 3)), ((4,), (3, 3)), ((4,), (3, 3))],
         [((2, 2), (3, 3)), ((1, 1, 1, 1), (3, 3)), ((2, 2), (3, 3))],
         [
             ((2, 2, 2), (3, 3, 3)),
             ((1, 2, 2, 1), (2, 3, 4)),
-            ((1, 1, 1, 1, 1, 1), (2, 1, 2, 1, 3)),
+            ((1, 2, 2, 1), (2, 3, 4)),
         ],
         [((1, np.nan), (3, 3)), ((1, np.nan), (2, 2, 2)), ((1, np.nan), (2, 1, 1, 2))],
-        [((4,), (1, 1, 1)), ((1, 1, 1, 1), (3,)), ((4,), (1, 1, 1))],
+        [((4,), (1, 1, 1)), ((1, 1, 1, 1), (3,)), ((4,), (3,))],
     ],
 )
 def test_calculate_prechunking_2d(old, new, expected):
-    actual = _calculate_prechunking(old, new)
+    actual = _calculate_prechunking(old, new, np.dtype(np.int16), None)
+    assert actual == expected
+
+
+@pytest.mark.parametrize(
+    ["old", "new", "expected"],
+    [
+        (
+            ((2, 2), (1, 1, 1, 1), (1, 1, 1, 1)),
+            ((1, 1, 1, 1), (4,), (2, 2)),
+            ((2, 2), (4,), (1, 1, 1, 1)),
+        ),
+        (
+            ((2, 2), (1, 1, 1, 1), (1, 1, 1, 1)),
+            ((1, 1, 1, 1), (2, 2), (2, 2)),
+            ((2, 2), (2, 2), (2, 2)),
+        ),
+        (
+            ((2, 2), (1, 1, 1, 1), (1, 1, 1, 1)),
+            ((1, 1, 1, 1), (2, 2), (4,)),
+            ((2, 2), (2, 2), (2, 2)),
+        ),
+        (
+            ((1, 1, 1, 1), (1, 1, 1, 1), (2, 2)),
+            ((2, 2), (4,), (1, 1, 1, 1)),
+            ((2, 2), (2, 2), (2, 2)),
+        ),
+    ],
+)
+def test_calculate_prechunking_3d(old, new, expected):
+    with dask.config.set({"array.chunk-size": "16 B"}):
+        actual = _calculate_prechunking(old, new, np.dtype(np.int16), None)
+    assert actual == expected
+
+
+@pytest.mark.parametrize(
+    ["chunk_size", "expected"],
+    [
+        ("1 B", ((10,), (1,) * 10)),
+        ("20 B", ((10,), (1,) * 10)),
+        ("40 B", ((10,), (2, 2, 1, 2, 2, 1))),
+        ("100 B", ((10,), (5, 5))),
+    ],
+)
+def test_calculate_prechunking_concatenation(chunk_size, expected):
+    old = ((10,), (1,) * 10)
+    new = ((2,) * 5, (5, 5))
+    with dask.config.set({"array.chunk-size": chunk_size}):
+        actual = _calculate_prechunking(old, new, np.dtype(np.int16), None)
+    assert actual == expected
+
+
+def test_calculate_prechunking_does_not_concatenate_object_type():
+    old = ((10,), (1,) * 10)
+    new = ((2,) * 5, (5, 5))
+
+    # Ensure that int dtypes get concatenated
+    new = ((2,) * 5, (5, 5))
+    with dask.config.set({"array.chunk-size": "100 B"}):
+        actual = _calculate_prechunking(old, new, np.dtype(np.int16), None)
+    assert actual == ((10,), (5, 5))
+
+    # Ensure object dtype chunks do not get concatenated
+    with dask.config.set({"array.chunk-size": "100 B"}):
+        actual = _calculate_prechunking(old, new, np.dtype(object), None)
+    assert actual == old
+
+
+@pytest.mark.parametrize(
+    ["old", "new", "expected"],
+    [
+        [((2, 2), (3, 3)), ((4,), (3, 3)), ((2, 2), (3, 3))],
+        [
+            ((2, 2, 2), (3, 3, 3)),
+            ((1, 2, 2, 1), (2, 3, 4)),
+            ((1, 1, 1, 1, 1, 1), (2, 1, 2, 1, 3)),
+        ],
+        [((4,), (1, 1, 1)), ((1, 1, 1, 1), (3,)), ((4,), (1, 1, 1))],
+    ],
+)
+def test_calculate_prechunking_splitting(old, new, expected):
+    # _calculate_prechunking does not concatenate on object
+    actual = _calculate_prechunking(old, new, np.dtype(object), None)
     assert actual == expected