Fix provisioner issue for on-demand/coretime #3141
Comments
|
This should be fine and is generally the same as #664, Ever since we refactored this out of the scheduler and into the |
|
Having a deeper look at the code and speaking with @antonva about this, we came to the following conclusion: Assume core A is occupied by Para1, the So it's a slight inefficiency that is made less likely due to the core affinities. I suggest we fix it by modifying the provisioner to check if the newly scheduled para is different than the previous one and if it is, request a backable candidate with the required_path of the new para instead (taking into account a possible parent candidate for Para 2 present on another core) |
|
I found a very similar problem in collation-generation: #3327 |
#3130 builds on top of #3160 Processes the availability cores and builds a record of how many candidates it should request from prospective-parachains and their predecessors. Tries to supply as many candidates as the runtime can back. Note that the runtime changes to back multiple candidates per para are not yet done, but this paves the way for it. The following backing/inclusion policy is assumed: 1. the runtime will never back candidates of the same para which don't form a chain with the already backed candidates. Even if the others are still pending availability. We're optimistic that they won't time out and we don't want to back parachain forks (as the complexity would be huge). 2. if a candidate is timed out of the core before being included, all of its successors occupying a core will be evicted. 3. only the candidates which are made available and form a chain starting from the on-chain para head may be included/enacted and cleared from the cores. In other words, if para head is at A and the cores are occupied by B->C->D, and B and D are made available, only B will be included and its core cleared. C and D will remain on the cores awaiting for C to be made available or timed out. As point (2) above already says, if C is timed out, D will also be dropped. 4. The runtime will deduplicate candidates which form a cycle. For example if the provisioner supplies candidates A->B->A, the runtime will only back A (as the state output will be the same) Note that if a candidate is timed out, we don't guarantee that in the next relay chain block the block author will be able to fill all of the timed out cores of the para. That increases complexity by a lot. Instead, the provisioner will supply N candidates where N is the number of candidates timed out, but doesn't include their successors which will be also deleted by the runtime. This'll be backfilled in the next relay chain block. Adjacent changes: - Also fixes: #3141 - For non prospective-parachains, don't supply multiple candidates per para (we can't have elastic scaling without prospective parachains enabled). paras_inherent should already sanitise this input but it's more efficient this way. Note: all of these changes are backwards-compatible with the non-elastic-scaling scenario (one core per para).
github-project-automation
bot
moved this from In Progress
to Completed
in parachains team board
…ch#3233) paritytech#3130 builds on top of paritytech#3160 Processes the availability cores and builds a record of how many candidates it should request from prospective-parachains and their predecessors. Tries to supply as many candidates as the runtime can back. Note that the runtime changes to back multiple candidates per para are not yet done, but this paves the way for it. The following backing/inclusion policy is assumed: 1. the runtime will never back candidates of the same para which don't form a chain with the already backed candidates. Even if the others are still pending availability. We're optimistic that they won't time out and we don't want to back parachain forks (as the complexity would be huge). 2. if a candidate is timed out of the core before being included, all of its successors occupying a core will be evicted. 3. only the candidates which are made available and form a chain starting from the on-chain para head may be included/enacted and cleared from the cores. In other words, if para head is at A and the cores are occupied by B->C->D, and B and D are made available, only B will be included and its core cleared. C and D will remain on the cores awaiting for C to be made available or timed out. As point (2) above already says, if C is timed out, D will also be dropped. 4. The runtime will deduplicate candidates which form a cycle. For example if the provisioner supplies candidates A->B->A, the runtime will only back A (as the state output will be the same) Note that if a candidate is timed out, we don't guarantee that in the next relay chain block the block author will be able to fill all of the timed out cores of the para. That increases complexity by a lot. Instead, the provisioner will supply N candidates where N is the number of candidates timed out, but doesn't include their successors which will be also deleted by the runtime. This'll be backfilled in the next relay chain block. Adjacent changes: - Also fixes: paritytech#3141 - For non prospective-parachains, don't supply multiple candidates per para (we can't have elastic scaling without prospective parachains enabled). paras_inherent should already sanitise this input but it's more efficient this way. Note: all of these changes are backwards-compatible with the non-elastic-scaling scenario (one core per para).
Provisioner code does not seem to be fully on-demand ready.
polkadot-sdk/polkadot/node/core/provisioner/src/lib.rs
Line 662 in a02b534
The text was updated successfully, but these errors were encountered: