Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

fix(ai): update ai-video selection suspension #3033

Open
wants to merge 6 commits into
base: ai-video
Choose a base branch
from

Conversation

ad-astra-video
Copy link
Contributor

@ad-astra-video ad-astra-video commented Apr 28, 2024

What does this pull request do? Explain your changes. (required)
Draft of ai-video selection algo fix.

Suspension was not working because the penalty was always 3. This logic was a carryover from transcoding where the suspender always started at a refresh count of 0 because a new session manager was created with each stream. For AI, we are reusing the session manager and the suspender so the refresh count does not reset between requests. The fix to suspension is to consider the current refresh count when calculating the penalty so it is 3 more than the current refresh count in the suspender.

There was also an issue where the discoveryPoolSize was always 100 and with limited orchestrators providing models a refresh of sessions was being done with every request. I added an initialPoolSize field to track the last refresh pool size to use with the shouldRefreshSessions logic rather than 100. This stabilizes the suspender to allow more orchestrators to be tried with each Select call.

Last update was moving the signalRefresh() for the suspender that increments the refresh counter in the suspender to the Refresh function makes it more stable that every time we refresh sessions we add to the suspender refresh count

Happy to segregate some of these changes to separate PRs. The suspension fixes can be added separately without dependency on ai-worker PR.

Specific updates (required)

  • Updates suspender to use the current refresh count of the suspender in the selector.
  • Moves penalty to the AISessionSelector to make it easier to update and available for calculations on the suspension needed
  • releases all Os when there are none in the warm and cold pool
  • Adds option to not use managed containers.

How did you test each of these updates (required)

I have been running these updates on my gateway. Tested 1-200 requests with 5-10 workers sending to gateway. All completed with 1-2 orchestrators providing Bytedance model.

Does this pull request close any open issues?

Checklist:

@github-actions github-actions bot added the AI Issues and PR related to the AI-video branch. label Apr 28, 2024
@ad-astra-video ad-astra-video force-pushed the ai-video-fix-selection-pr branch 2 times, most recently from 494b5d9 to 2504355 Compare May 7, 2024 11:24
@ad-astra-video ad-astra-video marked this pull request as ready for review July 22, 2024 12:19
@ad-astra-video ad-astra-video changed the title Ai video fix selection pr fix(ai): update ai-video selection suspension Aug 27, 2024
Copy link
Member

@victorges victorges left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Feel like I don't have context to officially approve this, but left some comments. Only nits tho, the implementation makes sense for the PR description.

// as well
// as well. Since AISessionManager re-uses the pools the suspension
// penalty needs to consider the current suspender count to set the penalty
last_count, ok := pool.suspender.list[sess.Transcoder()]
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit/lint: Vars in go should be camelCase

Comment on lines 234 to 235
// Refresh if the # of sessions across warm and cold pools falls below the smaller of the maxRefreshSessionsThreshold and
// 1/2 the total # of orchs that can be queried during discovery
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This comment seems out of place now, can you move it closer to L247?

@@ -222,7 +233,17 @@ func (sel *AISessionSelector) Select(ctx context.Context) *AISession {
shouldRefreshSelector := func() bool {
// Refresh if the # of sessions across warm and cold pools falls below the smaller of the maxRefreshSessionsThreshold and
// 1/2 the total # of orchs that can be queried during discovery
discoveryPoolSize := sel.node.OrchestratorPool.Size()
discoveryPoolSize := int(math.Min(float64(sel.node.OrchestratorPool.Size()), float64(sel.initialPoolSize)))
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why do we need this Min now? Can the pool grow from its initial size?

// penalty needs to consider the current suspender count to set the penalty
last_count, ok := pool.suspender.list[sess.Transcoder()]
if ok {
penalty = pool.suspender.count - last_count + pool.penalty
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm a little lost with this suspension logic. So, I see that:

  • the pool.suspender.count is increased every time signalRefresh() is called
  • pool.penalty is always set to 3
  • last_count is always set to suspender.count + 3

So, that logic would mean that we're not taking the suspended orchestrator until 3 times the signalRefresh() is called. Is this the idea of this suspension mechanism? That we don't allow the given O to get selected in the 3 refresh sessions?

// if there are no orchestrators in the pools
clog.Infof(ctx, "refreshing sessions, no orchestrators in pools")
for i := 0; i < sel.penalty; i++ {
sel.suspender.signalRefresh()
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

release all orchestrators

shouldn't we then just remove them from the suspender.list() rather than calling signalRefresh()? My understanding is that if penalty = 3, then we would need to call signalRefresh() 3 times in order to "release all orchestrators from suspension".

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
AI Issues and PR related to the AI-video branch.
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants