Reduce memory footprint of culling P2P rechunking (#8845)

Author: hendrikmakait

distributed/shuffle/_rechunk.py

@@ -370,28 +370,31 @@ def cull(
         indices_to_keep = self._keys_to_indices(keys)
         _old_to_new = old_to_new(self.chunks_input, self.chunks)
 
-        culled_deps: defaultdict[Key, set[Key]] = defaultdict(set)
-        for nindex in indices_to_keep:
-            old_indices_per_axis = []
-            keepmap[nindex] = True
-            for index, new_axis in zip(nindex, _old_to_new):
-                old_indices_per_axis.append(
-                    [old_chunk_index for old_chunk_index, _ in new_axis[index]]
-                )
-            for old_nindex in product(*old_indices_per_axis):
-                culled_deps[(self.name,) + nindex].add((self.name_input,) + old_nindex)
+        for ndindex in indices_to_keep:
+            keepmap[ndindex] = True
 
-        # Protect against mutations later on with frozenset
-        frozen_deps = {
-            output_task: frozenset(input_tasks)
-            for output_task, input_tasks in culled_deps.items()
-        }
+        culled_deps = {}
+        # Identify the individual partial rechunks
+        for ndpartial in _split_partials(_old_to_new):
+            # Cull partials for which we do not keep any output tasks
+            if not np.any(keepmap[ndpartial.new]):
+                continue
+
+            # Within partials, we have all-to-all communication.
+            # Thus, all output tasks share the same input tasks.
+            deps = frozenset(
+                (self.name_input,) + ndindex
+                for ndindex in _ndindices_of_slice(ndpartial.old)
+            )
+
+            for ndindex in _ndindices_of_slice(ndpartial.new):
+                culled_deps[(self.name,) + ndindex] = deps
 
         if np.array_equal(keepmap, self.keepmap):
-            return self, frozen_deps
+            return self, culled_deps
         else:
             culled_layer = self._cull(keepmap)
-            return culled_layer, frozen_deps
+            return culled_layer, culled_deps
 
     def _construct_graph(self) -> _T_LowLevelGraph:
         import numpy as np
@@ -695,14 +698,12 @@ def _slice_new_chunks_into_partials(
     return tuple(sliced_axes)
 
 
-def _partial_ndindex(ndslice: NDSlice) -> np.ndindex:
-    import numpy as np
-
-    return np.ndindex(tuple(slice.stop - slice.start for slice in ndslice))
+def _ndindices_of_slice(ndslice: NDSlice) -> Iterator[NDIndex]:
+    return product(*(range(slc.start, slc.stop) for slc in ndslice))
 
 
-def _global_index(partial_index: NDIndex, partial_offset: NDIndex) -> NDIndex:
-    return tuple(index + offset for index, offset in zip(partial_index, partial_offset))
+def _partial_index(global_index: NDIndex, partial_offset: NDIndex) -> NDIndex:
+    return tuple(index - offset for index, offset in zip(global_index, partial_offset))
 
 
 def partial_concatenate(
@@ -802,8 +803,8 @@ def partial_rechunk(
     )
 
     transfer_keys = []
-    for partial_index in _partial_ndindex(ndpartial.old):
-        global_index = _global_index(partial_index, old_partial_offset)
+    for global_index in _ndindices_of_slice(ndpartial.old):
+        partial_index = _partial_index(global_index, old_partial_offset)
 
         input_key = (input_name,) + global_index
 
@@ -822,8 +823,8 @@ def partial_rechunk(
     dsk[_barrier_key] = (shuffle_barrier, partial_token, transfer_keys)
 
     new_partial_offset = tuple(axis.start for axis in ndpartial.new)
-    for partial_index in _partial_ndindex(ndpartial.new):
-        global_index = _global_index(partial_index, new_partial_offset)
+    for global_index in _ndindices_of_slice(ndpartial.new):
+        partial_index = _partial_index(global_index, new_partial_offset)
         if keepmap[global_index]:
             dsk[(unpack_group,) + global_index] = (
                 rechunk_unpack,