Executor 2.0: Per-operator stream assignment policy #5620
Merged
Conversation
This file contains 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
mzient
added a commit
that referenced
this pull request
Sep 4, 2024
* Add ExecNode stream assignment algorithms and tests. NOTE: Follow up #5620 implements per-operator assignment. --------- Signed-off-by: Michal Zientkiewicz <michalz@nvidia.com>
mzient
force-pushed
the
per_operator_stream_assignment
branch
from
610fd7d
to
2a11ce6
Compare
|
CI MESSAGE: [18125248]: BUILD STARTED |
|
CI MESSAGE: [18125248]: BUILD PASSED |
szkarpinski
reviewed
Sep 6, 2024
| void Assign(ExecGraph &graph) { | ||
| int num_nodes = graph.Nodes().size(); | ||
|
|
||
| // the nodes in the graph must be sorted topologically |
There was a problem hiding this comment.
Such requirements that are only expressed in the comments make this bug-prone. How about adding a simple check that the graph is actually toposorted? It's just something like for each E in edges assert(e.from.index < e.to.index) - linear time, few lines of code
There was a problem hiding this comment.
I should rephrase: the graph is sorted. That's how it's built. You can manually construct one that's not, but it's not usable and its validation will fail.
It works by traversing the graph in depth-first fashion, from outputs to inputs. Each node has a set of "ready" streams which are merged when the graph converges. When a stream is reused in a subsequent node, the stream id is removed from the producer's ready set and the removal happens recursively, but only once for each stream id. Typical complexity is O(V+E), pessimistic complexity is O(V+E^2). Signed-off-by: Michal Zientkiewicz <michalz@nvidia.com>
Signed-off-by: Michal Zientkiewicz <michalz@nvidia.com>
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
mzient
force-pushed
the
per_operator_stream_assignment
branch
from
a80eb0b
to
0870fed
Compare
|
CI MESSAGE: [18189246]: BUILD STARTED |
szkarpinski
approved these changes
Sep 6, 2024
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [18190005]: BUILD STARTED |
stiepan
approved these changes
Sep 6, 2024
|
CI MESSAGE: [18190005]: BUILD PASSED |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Category:
New feature (non-breaking change which adds functionality)
Description:
The PR implements a per-operator stream assignment algorithm.
The algorithm minimizes the number of necessary streams and tries to minimize the number of stream switches.
Additional information:
Affected modules and functionalities:
Key points relevant for the review:
Tests:
Checklist
Documentation
DALI team only
Requirements
REQ IDs: N/A
JIRA TASK: N/A