Recommended production settings for Apache Cassandra 

On this page:

• Use the latest Java Virtual Machine
• Synchronize clocks
• TCP settings
• Disable CPU frequency scaling
• Make sure that new settings persist after reboot
• Optimize SSDs
• Use the optimum --setra setting for RAID on SSD
• Disable zone_reclaim_mode on NUMA systems
• Set user resource limits
• Disable swap
• Check the Java Hugepages setting
• Set the heap size for optional Java garbage collection in DataStax Enterprise
• Determining the heap size when using Concurrent-Mark-Sweep (CMS) garbage collection in DataStax Enterprise
• Set the heap size for optimal Java garbage collection in Cassandra
• Apply optimum blockdev --setra settings for RAID on spinning disks

Use the latest Java Virtual Machine 

Use the latest 64-bit version of Oracle Java Platform, Standard Edition 8 (JDK) or OpenJDK 8.

Synchronize clocks 

Synchronize the clocks on all nodes, using NTP (Network Time Protocol) or other methods.

This is required because Apache Cassandra™ overwrites a column only when there is another version whose timestamp is more recent, which can happen when machines in are different locations.

Cassandra timestamps are encoded as microseconds since UNIX epoch without timezone information. The timestamp for all writes in Cassandra is UTC (Universal Time Coordinated). DataStax recommends converting to local time only when generating output to be read by humans.

TCP settings 

To handle thousands of concurrent connections used by Cassandra, the following settings are recommended for optimizing the Linux network stack. Add these settings to /etc/sysctl.conf.

net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
net.core.rmem_default = 16777216
net.core.wmem_default = 16777216
net.core.optmem_max = 40960
net.ipv4.tcp_rmem = 4096 87380 16777216
net.ipv4.tcp_wmem = 4096 65536 16777216

$ sudo sysctl -p /etc/sysctl.conf

$ sudo sysctl -p /etc/sysctl.d/filename.conf

Disable CPU frequency scaling 

Recent Linux systems include a feature called CPU frequency scaling or CPU speed scaling. It allows a server's clock speed to be dynamically adjusted so that the server can run at lower clock speeds when the demand or load is low. This reduces the server's power consumption and heat output (which significantly impacts cooling costs). Unfortunately, this behavior has a detrimental effect on servers running Cassandra because throughput can get capped at a lower rate.

On most Linux systems, a CPUfreq governor manages the scaling of frequencies based on defined rules and the default ondemand governor switches the clock frequency to maximum when the demand is high and switches to the lowest frequency when the system is idle.

for CPUFREQ in /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
do
    [ -f $CPUFREQ ] || continue
    echo -n performance > $CPUFREQ
done

For more information, see High server load and latency when CPU frequency scaling is enabled in the DataStax Help Center.

Make sure that new settings persist after reboot 

Optimize SSDs 

The default SSD configurations on most Linux distributions are not optimal. Follow these steps to ensure the best settings for SSDs:

1. Ensure that the SysFS rotational flag is set to false (zero).

   This overrides any detection by the operating system to ensure the drive is considered an SSD.

2. Apply the same rotational flag setting for any block devices created from SSD storage, such as mdarrays.
3. Set the IO scheduler to either deadline or noop:
   • The noop scheduler is the right choice when the target block device is an array of SSDs behind a high-end IO controller that performs IO optimization.
   • The deadline scheduler optimizes requests to minimize IO latency. If in doubt, use the deadline scheduler.
4. Set the readahead value for the block device to 8 KB.

   This setting tells the operating system not to read extra bytes, which can increase IO time and pollute the cache with bytes that weren’t requested by the user.

   For example, if the SSD is /dev/sda, in /etc/rc.local:

   $ echo deadline > /sys/block/sda/queue/scheduler
   #OR...
   #echo noop > /sys/block/sda/queue/scheduler
   touch /var/lock/subsys/local
   echo 0 > /sys/class/block/sda/queue/rotational
   echo 8 > /sys/class/block/sda/queue/read_ahead_kb

Use the optimum --setra setting for RAID on SSD 

The optimum readahead setting for RAID on SSDs (in Amazon EC2) is 8KB, the same as it is for non-RAID SSDs. For details, see Optimizing SSDs.

Disable zone_reclaim_mode on NUMA systems 

$ echo 0 > /proc/sys/vm/zone_reclaim_mode

For more information, see Peculiar Linux kernel performance problem on NUMA systems on Apache Cassandra.

Set user resource limits 

Use the ulimit -a command to view the current limits. Although limits can also be temporarily set using this command, DataStax recommends making the changes permanent.

<cassandra_user> - memlock unlimited
<cassandra_user> - nofile 100000
<cassandra_user> - nproc 32768
<cassandra_user> - as unlimited

<cassandra_user> - memlock unlimited
<cassandra_user> - nofile 100000
<cassandra_user> - nproc 32768
<cassandra_user> - as unlimited

root - memlock unlimited
root - nofile 100000
root - nproc 32768
root - as unlimited

cassandra_user - nproc 32768

vm.max_map_count = 1048575

session    required   pam_limits.so

$ sudo sysctl -p

$ cat /proc/pid/limits

For more information, see Insufficient user resource limits errors.

Disable swap 

Failure to disable swap entirely can severely lower performance. Because Cassandra has multiple replicas and transparent failover, it is preferable for a replica to be killed immediately when memory is low rather than go into swap. This allows traffic to be immediately redirected to a functioning replica instead of continuing to hit the replica that has high latency due to swapping. If your system has a lot of DRAM, swapping still lowers performance significantly because the OS swaps out executable code so that more DRAM is available for caching disks.

If you insist on using swap, you can set vm.swappiness=1. This allows the kernel swap out the absolute least used parts.

$ sudo swapoff --all

To make this change permanent, remove all swap file entries from /etc/fstab.

For more information, see Nodes seem to freeze after some period of time.

Check the Java Hugepages setting

Many modern Linux distributions ship with Transparent Hugepages enabled by default. When Linux uses Transparent Hugepages, the kernel tries to allocate memory in large chunks (usually 2MB), rather than 4K. This can improve performance by reducing the number of pages the CPU must track. However, some applications still allocate memory based on 4K pages. This can cause noticeable performance problems when Linux tries to defrag 2MB pages. For more information, see Cassandra Java Huge Pages and this RedHat bug report.

$ echo never | sudo tee /sys/kernel/mm/transparent_hugepage/defrag

For more information, including a temporary fix, see No Apache Cassandra processing but high CPU usage.

Set the heap size for optional Java garbage collection in DataStax Enterprise 

Heap size is usually between ¼ and ½ of system memory. Do not devote all memory to heap because it is also used for offheap cache and file system cache.

If you don't see improved performance, contact the DataStax Services team for additional help in tuning the JVM.

Determining the heap size when using Concurrent-Mark-Sweep (CMS) garbage collection in DataStax Enterprise 

There are many nuances for tuning CMS. It requires time, expertise, and repeated testing to get the best results. DataStax recommends contacting the DataStax Services team instead. Tuning Java resources in Cassandra 2.1, or 3.0 provides the basic information to get you started.

Set the heap size for optimal Java garbage collection in Cassandra 

See Tuning Java resources in 2.1, 2.2, 3.0, or 3.x.

Apply optimum blockdev --setra settings for RAID on spinning disks 

Typically, a readahead of 128 is recommended.

Check to ensure setra is not set to 65536:

$ sudo blockdev --report /dev/spinning_disk

To set setra:

$ sudo blockdev --setra 128 /dev/spinning_disk