feat: [DSM-103] Improve scheduler fairness#9985
Open
alin-at-dfinity wants to merge 16 commits intomasterfrom
Open
feat: [DSM-103] Improve scheduler fairness#9985alin-at-dfinity wants to merge 16 commits intomasterfrom
alin-at-dfinity wants to merge 16 commits intomasterfrom
Conversation
Instead of a binary prioritized / opportunistic flag, explicitly (record and) prioritize long executions based on number of slices executed, AP and round when the long execution started. This ensures that we don't starve low priority canisters (which may happen with bounded AP and just the right distribution across execution cores). Also switch from persisting SubnetSchedule spread across individual canister states to persisting it as part of the subnet's SystemMetadata.
These are some scheduler improvements backported from the active canister scheduler dev branch, with the goal of having the scheduler efficiency proptest pass (with reduced success thresholds). * Fully segregate long and new executions across cores, to prevent inversion of priority when lower priority long executions get a slice executed. * Track long executions across iterations, not just the ones from the start of the round. * Charge heap delta rate limited canisters, so that they are actually skipped instead of just delayed while accumulating priority. * Charge immediately for the first round of every long execution (scheduled as a new execution). * Apply an exponential decay to AP outside the `[-2000, 500]` range (`-2000` because of 20 max DTS rounds, the `500` is somewhat arbitrary, but seems to work well). Due to the interaction between long and short executions, runaway priorities are inevitable. This provides a soft bound for runaway AP, while still preserving relative priorities to some extent. It also makes priority resets unnecessary (to be removed separately). * Fully distribute all positive free compute, even if it means exceeding 100 priority per canister (by distributing it equally to all canisters). * Allow long executions to use all scheduler cores if there are no new executions.
alin-at-dfinity
force-pushed
the
alin/DSM-103-scheduler-proptest
branch
from
b86e911 to
ac9fba3
Compare
alin-at-dfinity
changed the base branch from
master
to
alin/DSM-103-long-execution-priority-queue
alin-at-dfinity
changed the title
Base automatically changed from
alin/DSM-103-long-execution-priority-queue
to
master
April 27, 2026 09:53
… bail out before the latter if there are no active canisters.
…ction. And only do so once per round. * Tweak the per_canister_cap calculation so we always end up with sum(AP) >= 0. * Simplify long execution core calculation. * Improve some comments.
In field names and comments, replace "priority credit" and "executed slices" with "executed rounds". "Priority credit" was the old mechanism, now replaced. And "executed slices" is a misnomer, what we're actually counting is rounds during which a long execution made progress, not the actual number of slices executed (multiple slices might be executed in any given round, but we charge for rounds, not slices).
…ture::canister_priority_mut().
…harging for in-progress executions; and CanisterRoundState ordering.
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
1 participant
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 is a collection of scheduler improvements backported from the active canister scheduler dev branch, with the goal of having the scheduler efficiency proptest pass (with reduced success thresholds). And, by applying the proptest and all other
RoundScheduletests ontomaster, hugely reducing the size of that change.[-2000, 500]range (-2000because of 20 max DTS rounds, the500is somewhat arbitrary, but seems to work well). Due to the interaction between long and short executions, runaway priorities are inevitable. This provides a soft bound for runaway AP, while still preserving relative priorities to some extent. It also makes priority resets unnecessary (to be removed later).