Update: I check the source code: https://github.com/apache/cassandra/blob/4a0d1caa262af3b6f2b6d329e45766b4df845a88/src/java/org/apache/cassandra/tools/nodetool/Info.java#L154
Off heap memory usage comes from four parts. I rechecked cfstats to find out Bloom filter takes most of the off heap memory. I increased bloom_filter_fp_chance from the default 0.01 to 0.1 and restarted nodes. The memory usage comes down now. My table is kinda special in that it has lots of partitions but each partition is very small (maximum <1M). I guess for such a table, Bloom filter can be very large compared to the actual data size. On Fri, Jan 8, 2016 at 6:39 PM, Kai Wang <dep...@gmail.com> wrote: > Hi, > > When I switch a big table from STCS to LCS, I notice high off heap memory > usage using nodetool cfstats. "*Off heap memory used (total)*" shows +10G > usage. > > Eventually my nodes died because of OOM. How do I throttle off heap usage? > The only thing I see in cassandra.yaml is *memtable_offheap_space_in_mb. *Is > that the only knob? > > Thanks. >
