[Proposal] Enhance default routing and Placement Strategy using Directory Partitioning Scheme #9426
Comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
8000
{{ message }}
Uh oh!
There was an error while loading. Please reload this page.
Current client routing (sticky bucketing) and default grain placement strategies can cause unnecessary network hops, increasing activation latency and recovery churn, as they aren't optimally aligned with the grain's directory partition location.
This proposal has two parts. The first part is to change the default placement to
ResourceOptimizedPlacementand update that director so that it has 4 tiers instead of three:This optimizes for good locality between the caller & callee, and callee and directory, while still keeping the system balanced.
The second part of the idea capitalizes on the first by making clients route to the gateway corresponding to the grain's directory partition. It involves us removing the sticky bucketing system that we currently use and adding functionality to subscribe to membership changes from the cluster (eg, via an observer callback).
This should result in the majority of calls requiring no network round trips to the directory if they are submitted by an external client. Calls originating from a silo or grain can take advantage of locality (if within tolerance) and otherwise are co-located with their directory partition for improved activation perf (while still being much better balanced than
RandomPlacement, since the directory partitioning scheme has a very low chance of significant imbalance). It also reduces the grain directory recovery costs in the case of a silo crash.The text was updated successfully, but these errors were encountered: