Increased memory use on nodes after upgrading to DSE 5.0 or DSE 5.1

Summary

This article discusses an issue where nodes experience unusual performance issues after upgrading a cluster from DSE 4.8 to DSE 5.0 or DSE 5.1.

Applies to

• DataStax Enterprise 5.1
• DataStax Enterprise 5.0

Symptoms

Customers have reported performance issues not observed when clusters were running with DSE 4.8. Shortly after upgrading to DSE 5.0 or 5.1, nodes experience issues despite factors such as application load, cluster traffic, access patterns, time of day/week/month, and hardware/network resources being equal.

Symptoms include:

• Significant increase in off-heap memory usage (for example, when monitored using Linux utilities such as top or sar),
• Increased P99 and/or maximum read latencies (like from nodetool tablehistograms output)
• Intermittent read request timeouts in the extreme.

In a rare case, DSE being started on a node fails to finish the startup sequence because most of the memory gets consumed very quickly and is eventually exhausted so the Linux oom-killer terminates the DSE process.

Cause

Apache Cassandra uses memory-mapped file I/O through the Unix system call mmap() (or mmap for short). This mmap system call allows Cassandra to use the operating system's virtual memory to hold copies of data files so reading SSTables is fast. A hidden cassandra.yaml property called disk_access_mode determines how data files are accessed. The valid options are:

• auto (default) - both SSTable data and index files are mapped on 64-bit systems; only index files are mapped for 32-bit systems
• mmap - both data and index files are mapped to memory
• mmap_index_only - only index files are mapped to memory
• standard - Cassandra uses standard IO and no files are mapped to memory

In versions of Cassandra 2.1 or earlier (DataStax 4.8 or earlier), reading compressed SSTables involved the data being copied on-heap then sent to an on-heap buffer to be decompressed and behaved as if disk access mode was set to mmap_index_only despite the default mode being auto.

With the added support for direct buffer decompression in Cassandra 2.2 (CASSANDRA-8464), the behaviour for disk access mode changed to the way it was designed, i.e. default auto mode on 64-bit systems now mmap() both SSTable data and index files. Note that DSE 5.0 includes Cassandra 3.0 and DSE 5.1 includes Cassandra 3.11, see DataStax Enterprise, Apache Cassandra, CQL, and SSTable compatibility.

In cases where there are lots of random reads, and the set of SSTables being heavily read is larger than the available memory, the affected nodes will have a high number of page faults. In some cases, the affected servers run out of memory and the Linux oom-killer terminates DSE.

Solution

Since CASSANDRA-8464 allows mapping compressed data directly, it is more efficient to map only index files.

With the default disk_access_mode: auto during startup, DSE logs an entry similar to below:

INFO  [main] 2019-05-02 12:33:21,572  DatabaseDescriptor.java:350 - \
  DiskAccessMode 'auto' determined to be mmap, indexAccessMode is mmap

Set disk_access_mode to mmap_index_only in cassandra.yaml:

disk_access_mode: mmap_index_only

After restarting DSE, an entry in the logs will be similar to:

INFO [main] 2019-05-02 17:53:50,437 DatabaseDescriptor.java:356 - \
  DiskAccessMode is standard, indexAccessMode is mmap

This log entry indicates that SSTable data files are used with standard disk IO but index files will be mapped to memory.

WARNING - Enabling mmap in DSE 6 has a detrimental effect to a cluster's performance. See the article below for details.

See also

• KB article - Nodes become unresponsive during high app traffic periods when mmap is enabled on DSE 6
• DataStax Enterprise documentation for Cassandra.yaml

Credits

• Branimir Lambov
• Jake Luciani
• Mark Curtis
• Michael Keeney
• Stefania Alborghetti
• Thanh Tranh
• Wei Deng