[ https://issues.apache.org/jira/browse/CASSANDRA-19387?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Brad Schoening updated CASSANDRA-19387: --------------------------------------- Description: invalid was: I think we have been starving the OS of RAM on the micro 32GB hosts. Cassandra 4.0 switched to Netty (nio), which allocates Direct Memory buffers (off heap) and may have accentuated the issue. Reports are that Netty can be memory intensive [Netty 4.x High Memory Usage|https://access.redhat.com/solutions/5286701]. Some symptoms: * Summer 2023 RHEL begins swapping to favor page cache * Fall 2023 OutOfMemory DirectMemory seen on several clusters with 4.0.6 * Cluster 733 speed up 2X simply by switching from 32GB -> 64GB RAM Our typical configuration for Micro: 16GB JVM 16GB Direct Memory buffers for netty I/O (default is equal to heap size) --------- 0 memory left over for page cache or anything else DataStax DSE [recommends|https://docs.datastax.com/eol/en/dse/6.7/dse-admin/datastax_enterprise/config/configRecommendedSettings.html#configFileCacheSettings__chunkCacheHistory]: * Max direct memory = (system memory – JVM heap size) / 2 * Chunk cache = (Max direct memory ) / 2 For CaaS micro nodes, this suggests 16GB (heap) + 8GB (nio) + 2GB (chunk) + 0.5GB (networking) = 26.5 GB + CaaS agent (3GB?) = 29.5GB I’d like to see us use 48GB VMs for micro nodes. > Transition Micro config to 48GB RAM > ----------------------------------- > > Key: CASSANDRA-19387 > URL: https://issues.apache.org/jira/browse/CASSANDRA-19387 > Project: Cassandra > Issue Type: Improvement > Reporter: Brad Schoening > Priority: Normal > > invalid > -- This message was sent by Atlassian Jira (v8.20.10#820010) --------------------------------------------------------------------- To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org For additional commands, e-mail: commits-h...@cassandra.apache.org