Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
2.10
-
None
-
None
-
Docs Required, Release Notes Required
Description
There is a high threads contention for the intensive read/write scenario, caused by the PageMemoryNoStoreImpl rwLock.
When user tries to insert entries into a cache with a high (or better say very high) rate, when grid nodes are running on powerful machines (e.g. 96 cores, 390+ gigs of RAM each) then the JFR (Java Flight Recorder) report shows that like 30-40% of the time writing threads are waiting for the read lock for the page that they would like to modify (gray bards in the Threads tab). The issue seems to be worse when the cache has an associated SQL schema with indexes.
The workaround is to run multiple small grid nodes within the big host. However, this contradicts with the statement that Apache Ignite has a good support of both: horizontal and vertical scaling.