[ci] Remove hardcoded test shards#10743
Merged
areusch merged 1 commit intoapache:mainfrom Apr 8, 2022
driazati:generate_shard
Merged
[ci] Remove hardcoded test shards#10743areusch merged 1 commit intoapache:mainfrom driazati:generate_shard
areusch merged 1 commit intoapache:mainfrom
driazati:generate_shard
Conversation
areusch
reviewed
Apr 6, 2022
|
|
||
| {% macro sharded_test_step(name, num_shards, node, ws) %} | ||
| {% for shard_index in range(1, num_shards + 1) %} | ||
| '{{ name }} {{ shard_index }} of {{ num_shards }}': { |
Contributor
There was a problem hiding this comment.
minor nit: want to 0-pad shard_index and num_shards here?
Member
Author
There was a problem hiding this comment.
meh, this is intended for humans and everything is text-align: center-ed anyways so 0-padding won't make it easier to read IMO (also we are at like 3 shards max not 10 so we can revisit later)
Contributor
There was a problem hiding this comment.
ok sg. the main intent was that Jenkins sorts these alphabetically iiuc.
Member
Author
There was a problem hiding this comment.
ah I see, we should still be ok on that front but would probably need to pad if we shard more than n=9
This moves the sharding logic from being inlined in the Jenkinsfile to templated, so we can change just the number of shards and the test allocation in `conftest.py` and the Jenkinsfile will work to match. This also changes the test allocation from a manual balancing before to be random between shards. Each shard needs to know only its shard number and the total number of shards, then it hashes each test and skips it unless that hash falls within its allocated tests. This breaks up related tests across shards but has the downside that any change to the number of shards will shuffle around where the tests end up (but ideally this is rare as we settle on a good number of shards to use). This only does this for the GPU frontend tests but eventually we could expand it to more.
Member
Author
|
I also verified the tests ran in the various shards vs the tests from a recent main run, the overall result of tests ran is the same so the sharding is working correctly. Code is here https://gist.github.com/f636700cd68b5717350c107a0eaaee4e for the curious. |
areusch
approved these changes
Apr 8, 2022
|
|
||
| def pytest_collection_modifyitems(config, items): | ||
| if not all(k in os.environ for k in ["CI", "TVM_NUM_SHARDS", "TVM_SHARD_INDEX"]): | ||
| # Only apportion tests if in CI and in a job that is set up for it |
Contributor
There was a problem hiding this comment.
nit: could log here if CI is present
| "tests/python/topi/python/test_topi_conv2d_winograd.py::test_conv2d_nchw", | ||
| "tests/python/relay/test_py_converter.py::test_global_recursion", | ||
| ] | ||
| HARDCODED_ALLOCATIONS = {} |
Contributor
There was a problem hiding this comment.
nit: could do with dict comprehension: HARDCODED_ALLOCATIONS = {i: v for i, v in enumerate(_slowest_tests)}
mehrdadh
pushed a commit
to mehrdadh/tvm
that referenced
this pull request
Apr 11, 2022
This moves the sharding logic from being inlined in the Jenkinsfile to templated, so we can change just the number of shards and the test allocation in `conftest.py` and the Jenkinsfile will work to match. This also changes the test allocation from a manual balancing before to be random between shards. Each shard needs to know only its shard number and the total number of shards, then it hashes each test and skips it unless that hash falls within its allocated tests. This breaks up related tests across shards but has the downside that any change to the number of shards will shuffle around where the tests end up (but ideally this is rare as we settle on a good number of shards to use). This only does this for the GPU frontend tests but eventually we could expand it to more. Co-authored-by: driazati <driazati@users.noreply.github.com>
altanh
pushed a commit
to altanh/tvm
that referenced
this pull request
Apr 28, 2022
This moves the sharding logic from being inlined in the Jenkinsfile to templated, so we can change just the number of shards and the test allocation in `conftest.py` and the Jenkinsfile will work to match. This also changes the test allocation from a manual balancing before to be random between shards. Each shard needs to know only its shard number and the total number of shards, then it hashes each test and skips it unless that hash falls within its allocated tests. This breaks up related tests across shards but has the downside that any change to the number of shards will shuffle around where the tests end up (but ideally this is rare as we settle on a good number of shards to use). This only does this for the GPU frontend tests but eventually we could expand it to more. Co-authored-by: driazati <driazati@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This moves the sharding logic from being inlined in the Jenkinsfile to templated, so we can change just the number of shards and the test allocation in
conftest.pyand the Jenkinsfile will work to match. This also changes the test allocation from a manual balancing before to be random between shards. Each shard needs to know only its shard number and the total number of shards, then it hashes each test and skips it unless that hash falls within its allocated tests. This breaks up related tests across shards but has the downside that any change to the number of shards will shuffle around where the tests end up (but ideally this is rare as we settle on a good number of shards to use). Some tests are also manually allocated via round-robin to different shards to ensure that long-running tests run on different shards as much as possible.This only does this for the GPU frontend tests but eventually we could expand it to more.
cc @areusch