Hacker News Reader

This is exactly how the JVM CMS works, then what happens is the old gen becomes too fragmented to allocate, you end up with a GC failure and the JVM will rewrite the entire old heap.

When hadoop depended on a single namenode, a 64gb+ heap would take 5+ minutes to collect on these failure modes. Quite painful.

View on HN

Yes, but there are two easy solutions: 1. Larger nurseries to minimize churn in older generations. Fragmentation is a symptom of too many "false survivors". 2. Concurrent heap compaction.

The failure mode here is usually caching, in my experience. Lots of data, all dying old.

When JVMs started hitting a wall around 1GB of memory, we ended up moving long-lived data out of the old generation and into in-memory databases.

Reply on news.ycombinator.com