Adaptive queue for staging dials #237
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
This patch introduces an adaptive dial queue that spawns a dynamically sized set of goroutines to preemptively stage dials for later handoff to the DHT protocol for RPC. It identifies backpressure on both ends (dial consumers and dial producers), and takes compensating action by adjusting the worker pool. We start with `DialQueueMinParallelism` number of workers (6), and scale up and down based on demand and supply of dialled peers. The following events trigger scaling: - we scale up when we can't immediately return a successful dial to a new consumer. - we scale down when we've been idle for a while waiting for new dial attempts. - we scale down when we complete a dial and realise nobody was waiting for it. Dialler throttling (e.g. FD limit exceeded) is a concern, as we can easily spin up more workers to compensate, and end up adding fuel to the fire. Since we have no deterministic way to detect this for now, we hard-limit concurrency to `DialQueueMaxParallelism` (20).
ghost
assigned 
|
Future optimisation: cancelling pending dials to worse nodes as we find closer nodes to the target. EDIT: in practice, this is complex, because those theoretically better nodes may never respond, and we would've stopped making progress. The algorithm would have to compensate by backtracking and replaying those dials. Quite a dance. |
Stebalien
reviewed
Jan 29, 2019
|
Addressed the review comments, but I noticed a flaky test on CI along the way. I do deplore depending on time, but I cannot think of another way to test this. |
Stebalien
requested changes
Jan 29, 2019
Stebalien
requested changes
Jan 29, 2019
raulk
force-pushed
the
feat/adaptive-dial-queue
branch
from
4ab14f9
to
74d22f3
Compare
|
@Stebalien – up for re-review. I ended up changing the waiting mechanism to a slice, like we discussed in comments. |
Stebalien
approved these changes
Jan 30, 2019
4 tasks
This was referenced Jan 31, 2019
This was referenced Feb 8, 2019
Closed
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.
Currently the DHT is performing dials outside of the Alpha concurrency limit. We are dialling all nodes that peers return in
CloserPeerswithout limit. As a result, we end up flooding the swarm with dial jobs, which trips over the file descriptor limits, and brings dialling to a halt under some circumstances. Our current approach is also algorithmically incorrect, and leads to suboptimal query patterns.This patch introduces an adaptive dial queue that spawns a dynamically sized set of goroutines to preemptively stage dials for later handoff to the DHT protocol for RPC. It identifies backpressure on both ends (dial consumers and dial producers), and takes compensating action by adjusting the worker pool.
We start with
DialQueueMinParallelismnumber of workers (6), and scale up and down based on demand and supply of dialled peers.The following events trigger scaling:
Dialler throttling (e.g. FD limit exceeded) is a concern, as we can easily spin up more workers to compensate, and end up adding fuel to the fire. Since we have no deterministic way to detect this for now, we hard-limit concurrency to
DialQueueMaxParallelism(20).Testing this patch in a production mirror reduced dial backlog considerably, and showed the adaptiveness in action: