William Burns created ISPN-5308: ----------------------------------- Summary: Reduce ConcurrentHashMapv8 allocations Key: ISPN-5308 URL: https://issues.jboss.org/browse/ISPN-5308 Project: Infinispan Issue Type: Enhancement Components: Eviction Affects Versions: 7.2.0.Beta1 Reporter: William Burns Assignee: William Burns The current EquivalentConcurrentHashMapv8 allocates a lot of extra entries for the sake of atomicity. It would be helpful to reduce these allocations if possible. # Whenever a get refreshes an entries recency we have to allocate a new DequeNode. This is because of ConcurrentLinkedDeque not always removing the node (this can happen when the node is a head or tail node. If we are able to detect when the node was removed and reuse it it will drastically reduce how many allocations this uses since a deque should rarely be removed from the head or tail on a get. # The current size and current number of evicting elements is stored as 2 longs in a custom class. We have to use compare and set semantics to keep these correct. If instead we utilized a single long where the top 16 bits are reserved for number of evicting elements (concurrent puts) and used the bottom 48 bits for the actual size we can use a simple AtomicLong instead. -- This message was sent by Atlassian JIRA (v6.3.11#6341)