[GraphBolt] modify logic for `HeteroItemSet` indexing #7428

Skeleton003 · 2024-05-27T02:53:48Z

Description

First let's take a look at the current code for indexing a HeteroItemSet (occurs in HeteroItemSet.__getitem__):

        elif isinstance(index, Iterable):
            if not isinstance(index, torch.Tensor):
                index = torch.tensor(index)
            assert torch.all((index >= 0) & (index < self._length))
            key_indices = (
                torch.searchsorted(self._offsets, index, right=True) - 1
            )
            data = {}
            for key_id, key in enumerate(self._keys):
                mask = (key_indices == key_id).nonzero().squeeze(1)
                if len(mask) == 0:
                    continue
                data[key] = self._itemsets[key][
                    index[mask] - self._offsets[key_id]
                ]
            return data

Say the length of indices is N and the number of etypes/ntyeps is K, then the time complexity of current implementation of indexing a dictionary is O(N * K), which is mainly introduced by the line

mask = (key_indices == key_id).nonzero().squeeze(1)

If there are a lot of etypes, this line could easily become the bottleneck.

This draft PR intends to propose an alternative to current logic:

        elif isinstance(index, Iterable):
            if not isinstance(index, torch.Tensor):
                index = torch.tensor(index)
            sorted_index, indices = index.sort()
            assert sorted_index[0] >= 0 and sorted_index[-1] < self._length
            index_offsets = torch.searchsorted(sorted_index, self._offsets)
            data = {}
            for key_id, key in enumerate(self._keys):
                if index_offsets[key_id] == index_offsets[key_id + 1]:
                    continue
                current_indices, _ = indices[
                    index_offsets[key_id] : index_offsets[key_id + 1]
                ].sort()
                data[key] = self._itemsets[key][
                    index[current_indices] - self._offsets[key_id]
                ]
            return data

whose time complexity is O(N * logN) where the log is introduced by the sorting operation.

This will imporve the performance when there are many etypes, but might cause more time consuming when there are few etypes. A thoughtful consideration lies in striking a balance between the two approaches.

Update on June 18

Benchmark: https://docs.google.com/document/d/1Bbmp8gMekiGIYYxEMVbmXSANRZlZ_nTNbhpWul4RaKA/edit?usp=sharing

The results show that the original algorithm is faster than the new algorithm (theoretical time complexity N*logN) for almost all batch_size and num_types.

Checklist

Please feel free to remove inapplicable items for your PR.

The PR title starts with [$CATEGORY] (such as [NN], [Model], [Doc], [Feature]])
I've leverage the tools to beautify the python and c++ code.
The PR is complete and small, read the Google eng practice (CL equals to PR) to understand more about small PR. In DGL, we consider PRs with less than 200 lines of core code change are small (example, test and documentation could be exempted).
All changes have test coverage
Code is well-documented
To the best of my knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change
Related issue is referred in this PR
If the PR is for a new model/paper, I've updated the example index here.

Changes

dgl-bot · 2024-05-27T02:54:17Z

To trigger regression tests:

@dgl-bot run [instance-type] [which tests] [compare-with-branch];
For example: @dgl-bot run g4dn.4xlarge all dmlc/master or @dgl-bot run c5.9xlarge kernel,api dmlc/master

dgl-bot · 2024-05-27T03:30:05Z

Commit ID: 6c3a7f2

Build ID: 1

Status: ✅ CI test succeeded.

Report path: link

Full logs path: link

Skeleton003 · 2024-06-13T05:09:41Z

@dgl-bot

dgl-bot · 2024-06-13T05:45:58Z

Commit ID: b4c0ea46dfefd4426afc83d6a6098b4495a653b5

Build ID: 2

Status: ✅ CI test succeeded.

Report path: link

Full logs path: link

mfbalin · 2024-06-17T03:09:07Z

Do you have a benchmark comparing the new approach to the old one for different K values?

mfbalin · 2024-06-17T03:09:51Z

python/dgl/graphbolt/itemset.py

-                data[key] = self._itemsets[key][
-                    index[mask] - self._offsets[key_id]
-                ]
+            if len(index) < self._threshold:


I think we need a benchmark before settings such a threshold only by looking at runtime complexity.

You're right. I'll do it right away.

Let's use K values 1 2 4 8 16 32 etc.

This is a little bit complex, I think the optimal complexity should be O(N*logK), with additional 2 helper array: offsets = [0, 10, 30, 60], etypes = ["A", "B", "C", "D"]

NVM, I see the points, to do the way I suggested, we need to implement our own C++ kernel with parallel optimization.

Is any of the ops used in the new implementation single threaded?

Seems no. Maybe sorting op itself is a bit slow even it's multi-threaded?

How did you verify?

I'm not sure, I'm just guessing. Did you find out anything from the benchmarking code?

dgl-bot · 2024-06-17T03:36:52Z

Commit ID: 50da718

Build ID: 3

Status: ❌ CI test failed in Stage [Distributed Torch CPU Unit test].

Report path: link

Full logs path: link

Skeleton003 · 2024-06-18T05:31:40Z

@mfbalin @frozenbugs See benchmark results in the description. The new implementation does not seem to be as efficient as we thought. Maybe we should keep it as is?

dgl-bot · 2024-06-18T06:08:04Z

Commit ID: 591c122323236759ec8e2df4021308331e93cf6b

Build ID: 4

Status: ✅ CI test succeeded.

Report path: link

Full logs path: link

mfbalin · 2024-06-18T13:36:44Z

@mfbalin @frozenbugs See benchmark results in the description. The new implementation does not seem to be as efficient as we thought. Maybe we should keep it as is?

Let me take a look at the code to see if we missed anything. Thank you for the benchmark.

mfbalin · 2024-06-28T00:12:58Z

python/dgl/graphbolt/itemset.py

+            if len(index) < self._threshold:
+                # Say N = len(index), and K = num_types.
+                # If logN < K, we use the algo with time complexity O(N*logN).
+                sorted_index, indices = index.sort()


Can we try numpy.argsort here to get indices and index[indices] to get sorted_index? It looks like numpy might have a more efficient sorting implementation. When benchmarking, we should ensure that we have a recent version of numpy installed. It looks like numpy uses this efficient sorting implementation by intel: https://github.com/intel/x86-simd-sort

This is assuming that the sort is the bottleneck for this code.

Thanks for your info!

When benchmarking, we should ensure that we have a recent version of numpy installed.

How recent is the very version? Because it seems that we have just diabled numpy>=2.0.0 in #7479 .

numpy/numpy#22315
They added faster sort in this PR. Looks like the version number is 1.25 or later.
https://github.com/search?q=repo%3Anumpy%2Fnumpy%20%2322315&type=code

Let's use the latest 1.x version and see how the performance is.

Though the code changes were commited in numpy/numpy#22315 where the version is 1.25, this improvement was not officially announced until NumPy 2.0.0 Release Notes. Therefore, it is likely that they did not integrate the changes until version 2.0.0.

I plan to move that we offer full support for numpy>=2.0.0 at the Monday meeting, and perform the benchmark after we do so.

numpy>=2 is compatible with DGL. I think we can perform this benchmark. I just ran the graphbolt tests with numpy>=2 installed and all passed.

See https://docs.google.com/document/d/1Bbmp8gMekiGIYYxEMVbmXSANRZlZ_nTNbhpWul4RaKA/edit?usp=sharing . The results seem to have changed little.

Thank you for the updated numbers. I will profile the benchmark code and see if there is a potential improvement we can do.

mfbalin · 2024-06-28T15:11:50Z

python/dgl/graphbolt/itemset.py

+                        continue
+                    current_indices, _ = indices[
+                        index_offsets[key_id] : index_offsets[key_id + 1]
+                    ].sort()


We could use np.sort here as well.

Ubuntu added 6 commits May 24, 2024 09:15

1

4d91a36

lint

e5b84c3

modify itemsampler

5968610

cls

187c8b8

itemsampler no need to sort

6aaa74b

revert changes in itemsampler

6c3a7f2

Skeleton003 changed the title ~~Heteroitemset getitem iter~~ [GraphBolt] modify logic for HeteroItemSet indexing May 27, 2024

Skeleton003 and others added 3 commits June 13, 2024 05:21

Merge branch 'master' into heteroitemset_getitem_iter

e7ae261

threshold

d70401c

property

eb94470

Skeleton003 marked this pull request as ready for review June 12, 2024 22:02

Skeleton003 requested review from frozenbugs, Rhett-Ying and mfbalin June 12, 2024 22:03

Merge branch 'master' into heteroitemset_getitem_iter

50da718

mfbalin reviewed Jun 17, 2024

View reviewed changes

mfbalin reviewed Jun 28, 2024

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[GraphBolt] modify logic for `HeteroItemSet` indexing #7428

[GraphBolt] modify logic for `HeteroItemSet` indexing #7428

Skeleton003 commented May 27, 2024 •

edited

Loading

dgl-bot commented May 27, 2024

dgl-bot commented May 27, 2024

Skeleton003 commented Jun 13, 2024

dgl-bot commented Jun 13, 2024

mfbalin commented Jun 17, 2024 •

edited

Loading

mfbalin Jun 17, 2024 •

edited

Loading

Skeleton003 Jun 17, 2024

mfbalin Jun 17, 2024

frozenbugs Jun 17, 2024

frozenbugs Jun 18, 2024

mfbalin Jun 18, 2024

Skeleton003 Jun 19, 2024

mfbalin Jun 19, 2024 •

edited

Loading

Skeleton003 Jun 19, 2024 •

edited

Loading

dgl-bot commented Jun 17, 2024

Skeleton003 commented Jun 18, 2024

dgl-bot commented Jun 18, 2024

mfbalin commented Jun 18, 2024

mfbalin Jun 28, 2024 •

edited

Loading

mfbalin Jun 28, 2024

Skeleton003 Jun 29, 2024

mfbalin Jun 29, 2024

mfbalin Jun 29, 2024

Skeleton003 Jun 29, 2024

mfbalin Jul 1, 2024

Skeleton003 Jul 1, 2024

mfbalin Jul 1, 2024

mfbalin Jun 28, 2024

[GraphBolt] modify logic for HeteroItemSet indexing #7428

Are you sure you want to change the base?

[GraphBolt] modify logic for HeteroItemSet indexing #7428

Conversation

Skeleton003 commented May 27, 2024 • edited Loading

Description

Update on June 18

Checklist

Changes

dgl-bot commented May 27, 2024

dgl-bot commented May 27, 2024

Skeleton003 commented Jun 13, 2024

dgl-bot commented Jun 13, 2024

mfbalin commented Jun 17, 2024 • edited Loading

mfbalin Jun 17, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mfbalin Jun 19, 2024 • edited Loading

Choose a reason for hiding this comment

Skeleton003 Jun 19, 2024 • edited Loading

Choose a reason for hiding this comment

dgl-bot commented Jun 17, 2024

Skeleton003 commented Jun 18, 2024

dgl-bot commented Jun 18, 2024

mfbalin commented Jun 18, 2024

mfbalin Jun 28, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

[GraphBolt] modify logic for `HeteroItemSet` indexing #7428

[GraphBolt] modify logic for `HeteroItemSet` indexing #7428

Skeleton003 commented May 27, 2024 •

edited

Loading

mfbalin commented Jun 17, 2024 •

edited

Loading

mfbalin Jun 17, 2024 •

edited

Loading

mfbalin Jun 19, 2024 •

edited

Loading

Skeleton003 Jun 19, 2024 •

edited

Loading

mfbalin Jun 28, 2024 •

edited

Loading