Data store avoid graph re-walk #5660
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
dwsutherland
force-pushed
the
data-store-avoid-graph-rewalk
branch
3 times, most recently
from
f10aa7b
to
65cb9fc
Compare
|
Coverage should improve with a n=2 window test. |
dwsutherland
force-pushed
the
data-store-avoid-graph-rewalk
branch
4 times, most recently
from
99a86e8
to
f110f63
Compare
dwsutherland
force-pushed
the
data-store-avoid-graph-rewalk
branch
4 times, most recently
from
714f758
to
58dbf6e
Compare
|
I've been having a play with this branch with the n=1 window in I've been playing with the following diff in order to track which children/parents are being walked over: diff --git a/cylc/flow/data_store_mgr.py b/cylc/flow/data_store_mgr.py
index b6a2960fb..e78cd767c 100644
--- a/cylc/flow/data_store_mgr.py
+++ b/cylc/flow/data_store_mgr.py
@@ -732,6 +732,7 @@ class DataStoreMgr:
None
"""
+ print(f'# increment_graph_window({source_tokens.relative_id})')
# common refrences
active_id = source_tokens.id
diff --git a/cylc/flow/taskdef.py b/cylc/flow/taskdef.py
index a4ae681e2..d2517383b 100644
--- a/cylc/flow/taskdef.py
+++ b/cylc/flow/taskdef.py
@@ -36,6 +36,7 @@ if TYPE_CHECKING:
def generate_graph_children(tdef, point):
"""Determine graph children of this task at point."""
+ print(f'# generate_graph_children({point}/{tdef.name})')
graph_children = {}
for seq, dout in tdef.graph_children.items():
for output, downs in dout.items():
@@ -78,6 +79,7 @@ def generate_graph_parents(tdef, point, taskdefs):
Infer parents be reversing upstream triggers that lead to point/task.
"""
+ print(f'# generate_graph_parents({point}/{tdef.name})')
graph_parents = {}
for seq, triggers in tdef.graph_parents.items():
if not seq.is_valid(point):Using the graph from your example above with the default n=1 window: flow.cylc[scheduler]
cycle point format = CCYY
[scheduling]
runahead limit = P1
initial cycle point = 2020
[[graph]]
P1Y = """
a => b => c => d => e => f => g
b => h => i => f
g[-P1Y] => a
"""
[runtime]
[[root]]
script = sleep 2
[[b, c, d, e, f, i]]
[[a]]
[[g,h]]It looks like the algorithm is looking one edge further infront ( E.G. here's a snippet from E.G. in the above output, when the n=0 window was When the children of |
|
Moving onto the n=2 window (and dumping out the n-window after each call) with this diff: diff --git a/cylc/flow/data_store_mgr.py b/cylc/flow/data_store_mgr.py
index b6a2960fb..3a21bbbea 100644
--- a/cylc/flow/data_store_mgr.py
+++ b/cylc/flow/data_store_mgr.py
@@ -473,7 +473,7 @@ class DataStoreMgr:
self.updated_state_families = set()
# Update workflow state totals once more post delta application.
self.state_update_follow_on = False
- self.n_edge_distance = 1
+ self.n_edge_distance = 2
self.next_n_edge_distance = None
self.latest_state_tasks = {
state: deque(maxlen=LATEST_STATE_TASKS_QUEUE_SIZE)
@@ -732,6 +732,7 @@ class DataStoreMgr:
None
"""
+ print(f'# increment_graph_window({source_tokens.relative_id})')
# common refrences
active_id = source_tokens.id
@@ -1018,6 +1019,10 @@ class DataStoreMgr:
self.prune_trigger_nodes[active_id])
del self.prune_trigger_nodes[active_id]
+ task_proxies = self.data[self.workflow_id][TASK_PROXIES]
+ for tp_id in sorted(list(task_proxies) + list(self.added[TASK_PROXIES])):
+ print(f'$ {tp_id}')
+
def generate_edge(
self,
parent_tokens: Tokens,
diff --git a/cylc/flow/taskdef.py b/cylc/flow/taskdef.py
index a4ae681e2..d2517383b 100644
--- a/cylc/flow/taskdef.py
+++ b/cylc/flow/taskdef.py
@@ -36,6 +36,7 @@ if TYPE_CHECKING:
def generate_graph_children(tdef, point):
"""Determine graph children of this task at point."""
+ print(f'# generate_graph_children({point}/{tdef.name})')
graph_children = {}
for seq, dout in tdef.graph_children.items():
for output, downs in dout.items():
@@ -78,6 +79,7 @@ def generate_graph_parents(tdef, point, taskdefs):
Infer parents be reversing upstream triggers that lead to point/task.
"""
+ print(f'# generate_graph_parents({point}/{tdef.name})')
graph_parents = {}
for seq, triggers in tdef.graph_parents.items():
if not seq.is_valid(point):Here's the log: Looks good, gives the right result, some possible savings to be made avoiding duplicate |
|
There are duplicate I spotted a [very] small optimisation moving that return a little higher up in the code: --- a/cylc/flow/data_store_mgr.py
+++ b/cylc/flow/data_store_mgr.py
@@ -1108,11 +1108,15 @@ class DataStoreMgr:
Orphan tasks with no children return (True, None) respectively.
"""
+ tp_id = tokens.id
+ task_proxies = self.data[self.workflow_id][TASK_PROXIES]
+ if tp_id in task_proxies or tp_id in self.added[TASK_PROXIES]:
+ # ghost task already created
+ return
+
name = tokens['task']
point_string = tokens['cycle']
t_id = self.definition_id(name)
- tp_id = tokens.id
- task_proxies = self.data[self.workflow_id][TASK_PROXIES]
is_orphan = False
if name not in self.schd.config.taskdefs:
@@ -1120,8 +1124,6 @@ class DataStoreMgr:
if itask is None:
itask = self.schd.pool.get_task(point_string, name)
- if tp_id in task_proxies or tp_id in self.added[TASK_PROXIES]:
- return
if itask is None:
itask = TaskProxy( |
|
Tried out the new graphDepth field, the numbers came out as expected so this should close #5609 which will allow us to unlock the higher n-window views in the GUI! Is the graph window that the data store is currently configured to track available in the schema? That would be handy for the UI. |
diff --git a/cylc/flow/data_store_mgr.py b/cylc/flow/data_store_mgr.py
index 03c77eb94..a085f2f3f 100644
--- a/cylc/flow/data_store_mgr.py
+++ b/cylc/flow/data_store_mgr.py
@@ -754,6 +754,7 @@ class DataStoreMgr:
# walk filling, that may not have been the entire walk).
# If walk already completed, must have gone from non-active to active
# again.. So redo walk (as walk nodes may be pruned).
+ print(f'# Active ID: {active_id}')
if (
active_id not in all_walks
or active_id in self.n_window_completed_walks
@@ -899,6 +900,7 @@ class DataStoreMgr:
)
continue
+ print(f'# Generate Children/Parents of {node_tokens.relative_id}')
# Children/downstream nodes
# TODO: xtrigger is workflow_state edges too
# Reference set for workflow relations
diff --git a/cylc/flow/taskdef.py b/cylc/flow/taskdef.py
index a4ae681e2..daae255f4 100644
--- a/cylc/flow/taskdef.py
+++ b/cylc/flow/taskdef.py
@@ -15,6 +15,7 @@
# along with this program. If not, see <http://www.gnu.org/licenses/>.
"""Task definition."""
+import inspect
from collections import deque
from typing import TYPE_CHECKING
@@ -36,6 +37,7 @@ if TYPE_CHECKING:
def generate_graph_children(tdef, point):
"""Determine graph children of this task at point."""
+ print(f'# Children of {point}/{tdef.name} - Called by {inspect.currentframe().f_back.f_code.co_name}')
graph_children = {}
for seq, dout in tdef.graph_children.items():
for output, downs in dout.items():
@@ -78,6 +80,7 @@ def generate_graph_parents(tdef, point, taskdefs):
Infer parents be reversing upstream triggers that lead to point/task.
"""
+ print(f'# parents of {point}/{tdef.name} - Called by {inspect.currentframe().f_back.f_code.co_name}')
graph_parents = {}
for seq, triggers in tdef.graph_parents.items():
if not seq.is_valid(point):
As you've spotted, I've been relying on
I think we can add that as a workflow field (if it's not already). I've also thought of an optimization for the depth finder, so will put that up in the morning. |
dwsutherland
force-pushed
the
data-store-avoid-graph-rewalk
branch
from
58dbf6e
to
bee6937
Compare
|
I changed it to reuse the |
dwsutherland
force-pushed
the
data-store-avoid-graph-rewalk
branch
3 times, most recently
from
471ebc4
to
529c3ca
Compare
|
There were still some duplicate |
dwsutherland
force-pushed
the
data-store-avoid-graph-rewalk
branch
from
529c3ca
to
76aecbc
Compare
dwsutherland
pushed a commit
to dwsutherland/cylc-flow
that referenced
this pull request
Oct 31, 2023
* Sometimes tasks may appear in the n-window, then dissappear only to reappear at a later time. * This is a difficult case for the `increment_graph_window` to handle. * See cylc#5660
dwsutherland
force-pushed
the
data-store-avoid-graph-rewalk
branch
from
f68a63b
to
7ca4035
Compare
This one turned out to be an easy fix: (needed to flag depth to be recalculated after window size change) |
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
* Sometimes tasks may appear in the n-window, then dissappear only to reappear at a later time. * This is a difficult case for the `increment_graph_window` to handle. * See cylc#5660
Co-authored-by: Oliver Sanders <oliver.sanders@metoffice.gov.uk>
* Test the added, updated and pruned deltas created by increment_graph_window as the window changes. * Test window resizing with n=1,2,3
dwsutherland
force-pushed
the
data-store-avoid-graph-rewalk
branch
from
7ca4035
to
bc1a017
Compare
|
Been back through the test workflow above from n=1 to n=4, no problems! Re-ran the one which showed the longest-path issues before, perfect (and kinda pretty):
Tried a bunch of workflows from my cylc-src dir, the algorithm is now producing the correct results for all of them 🎉! |
|
Profiling results for the following workflow in simulation mode: [task parameters]
x = 1..100
y = 1..100
[scheduling]
cycling mode = integer
initial cycle point = 1
final cycle point = 1
[[graph]]
P1 = d[-P1] => a => b<x> => c<y> => d
Good positive change, increases functionality whilst improving performance 💯! |
oliver-sanders
approved these changes
Nov 2, 2023
There was a problem hiding this comment.
Thanks for all your hard work @dwsutherland, this looked like a tricky one!
MetRonnie
added a commit
to MetRonnie/cylc-flow
that referenced
this pull request
Nov 2, 2023
oliver-sanders
added a commit
that referenced
this pull request
Nov 2, 2023
Added changelog for #5660
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.
addresses #5485, #5609
At present we've resolved an efficiency problem at n=1 window size by restricting the graph walk to only parents or only children (i.e. 'c' 'p', or 'ccc' 'ppp' for n=3) .. However this excludes cousins, which is evident at n>1 (i.e. 'cp' or 'pc' at n=2)
This restriction was a cheap way to avoid graph re-walks.. #5475
Here the problem is actually fixed at any window size, by changing the graph walk algorithm from a recursive to a walk location approach.. From the method's description:
Window resizing will now happen as soon as it's set (not as the graph progresses)
N-window edge depth of nodes is now found/updated with new field
graph_depth.. A good example workflow:Gives:
Performance
Taken from #5435
$ cylc play <id> --profile --no-detachRun this for a few mins, when you get bored ctrl+c it, then:
This PR is even faster than current (#5475) at
n=1:At
n=2there's a bit more load:but clearly not squared or cubed.
Check List
CONTRIBUTING.mdand added my name as a Code Contributor.setup.cfg(andconda-environment.ymlif present).CHANGES.mdentry included if this is a change that can affect users