[IGNITE-8681] Using ExpiryPolicy with persistence causes significant slowdown.

• Type: Improvement
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: 2.7
• Component/s: cache
• Labels: performance
• Assignee: Andrey Mashenkov
• Reporter: Andrey Mashenkov
• Created: 01/Jun/18 14:24
• Updated: 12/Dec/18 14:37
• Resolved: 03/Jul/18 16:21

Description

Almost all ignite operations calls CU.unwindEvicts() on finish to help to evict expired entries.
In unwindEvicts(), threads iterate over all node partitions and check every partition PendingTree for expired entries. This costs too much.

We already have a flag on per-cachegroup basis that indicated ExpiryPolicy is used. It raised once expiring entry has been put to cache or we initialize non-empty pending tree from persistence.
So, we have to optimize a case when there is no expired pending entries, but pending tree is non-empty.

We can add some throttling on per-partition basis to reduce useless pending tree lookups.
E.g. if there is nothing to clean, no thread should check partition during next 100ms interval.

Issue Links

• incorporates: IGNITE-8747 Remove\RemoveAll method should not count expired entry as removed.
• is caused by: IGNITE-5874 Store TTL expire times in B+ tree on per-partition basis
• is related to: IGNITE-8483 Per partition TTL pending tree for in-memory caches.
• links to: GitHub Pull Request #4110
• links to: TC run results