Jump to: navigation, search

Performance Tuning

Besides configuring keyspaces and column families, it is possible to further tweak the performance of Cassandra by editing cassandra.yaml (Node and Cluster Configuration) or by editing cassandra-env.sh (JVM Configuration).

Descriptions of tunable properties can be found in both cassandra.yaml and cassandra-env.sh. A summary of these properties can be seen in the tables below:

Performance Tuning Properties (cassandra.yaml)

Option

Default

Description

column_index_size_in_kb

64

The size at which column indexes are added to a row. Value should be kept small if only a select few columns are consistently read from each row as a higher value implies that more row data must be deserialized for each read (until index is added).

commitlog_sync

periodic

Allowed values are periodic or batch. In periodic mode, the value of commitlog_sync_period_in_ms determines how frequently the commitlog is synchronized to disk. Writes are acknowledged at every periodic sync. If set to batch, writes are not acknowledged until fsynced to disk.

commitlog_sync_period_in_ms

10000 (10 seconds)

Determines how often (in milliseconds) to sync commitlog to disk when commitlog_sync is set to periodic.

commitlog_total_space_in_mb

4096

When commitlog reaches specified size, Cassandra flushes memtables to disk for oldest commitlog segments. Reduces amount of data to replay on startup.

compaction_throughput_mb_per_sec

16

Throttles compaction to the given total throughput across entire system. Value should be proportional to rate of write throughput (16 to 32 times). Setting to 0 disables compaction throttling.

concurrent_compactors

1 (per CPU core)

Max number of concurrent compaction processes allowed on a node.

concurrent_reads

16

Recommended setting is 16 * number_of_drives. This allows enough operations to queue such that the OS and drives can reorder them and minimize disk fetches.

concurrent_writes

32

Number of concurrent writes should be proportional to number of CPU cores in system. Recommended setting is (8 * number_of_cpu_cores).


memtable_flush_writers

1 per data directory

Number of memtable flush writer threads. Influences flush performance and can be increased if you have a large Java heap size and many data directories.

memtable_heap_space_in_mb

1/4 of heap

Total memory used for all column family memtables on a node.

stream_throughput_outbound_megabits_per_sec

400

Max outbound throughput on a node for streaming file transfers.

JVM Configuration Settings

Linux: conf/cassandra-env.sh
Windows: bin\cassandra.bat

Option

Default

Description

MAX_HEAP_SIZE

Half of available physical memory

Maximum heap size for the JVM. Same value is used for minimum heap size, allowing heap to be locked in memory. Should be set in conjunction with HEAP_NEWSIZE.

HEAP_NEWSIZE

100 MB per physical CPU core

Size of young generation. Larger value leads to longer GC pause times while smaller value will typically lead to more expensive GC. Set in conjunction with MAX_HEAP_SIZE.

com.sun.management.jmxremote.port

7199

Port on which Cassandra listens for JMX connections.

com.sun.management.jmxremote.ssl

false

Enable/disable SSL for JMX.

com.sun.management.jmxremote.authenticate

false

Enable/disable remote authentication for JMX.


This page was last edited on March 31, 2016, at 22:22.
Comments or questions about this documentation? Contact us for support!