Memory and Disk Space

Minimum Memory

For a production environment, Hewlett Packard Enterprise recommends at least 32 GB of memory per node. The absolute memory requirement for each node is determined by the data-fabric services that are configured to run on the node. For each configured service, you can adjust the minimum and maximum memory or use the default values, depending on your performance and functionality requirements (see Allocating Memory for Nodes). Typical HPE Ezmeral Data Fabric production nodes have 128 GB or more. Development nodes often use considerably less than 32 GB.

Run free -g to display total and available memory in gigabytes.

$ free -g
              total        used        free      shared      buffers      cached
Mem:              3           2           1           0            0           1
-/+ buffers/cache:            0           2
Swap:             2           0           2

If the free command is not found, you can use other options such as grep MemTotal: /proc/meminfo, vmstat -s -SM, top, or various GUI system-information tools.

HPE does not recommend using the numad service, since it has not been tested and validated with HPE Ezmeral Data Fabric. Using the numad service can cause artificial memory constraints to be set, which can lead to performance degradation under load. To stop and disable the numad service:

Stop the service: systemctl stop numad.
Set the numad service not to start on reboot: systemctl disable numad

HPE does not recommend using always overcommit as it can lead to the kernel memory manager stopping processes to free memory, resulting in stopped HPE Ezmeral Data Fabric processes and system instability. Leave vm.overcommit_memory at its default value of 0, do not change the value to 1 or 2.

You can explore the functionality of HPE Ezmeral Data Fabric on non-production equipment, but under the demands of a production environment, memory needs to be balanced against disks, network, and CPU.

Storage

For data disks, Installer versions 1.12.0.0 and later require a minimum disk size that is equal to the physical memory on the node. If a data disk does not meet the minimum disk size requirement, a verification error is generated.

To display the currently available disks, use a command such as the following:

ls -l /dev/sd*
brw-rw---- 1 root 1000 8,  0 Sep 14 23:49 /dev/sda
brw-rw---- 1 root disk 8,  1 Sep 14 23:49 /dev/sda1
brw-rw---- 1 root disk 8,  2 Sep 14 23:49 /dev/sda2
brw-rw---- 1 root mapr 8, 16 Sep 20 11:44 /dev/sdb
brw-rw---- 1 root mapr 8, 32 Sep 20 11:44 /dev/sdc
brw-rw---- 1 root mapr 8, 48 Sep 20 11:44 /dev/sdd

To check the available disk space:

df /dev/sda
Filesystem     1K-blocks  Used Available Use% Mounted on
devtmpfs        12225720     0  12225720   0% /dev

HPE Ezmeral Data Fabric software works with raw unformatted devices and partitions. For optimized performance and high reliability, HPE recommends using raw unformatted devices. For data nodes, allocate at least three unmounted physical drives or partitions for data-fabric storage. Data Fabric software uses disk spindles in parallel for faster read/write bandwidth and therefore groups disks into sets of three.

Minimum Disk Allocation: HPE Ezmeral Data Fabric software requires a minimum of one disk or partition for data-fabric data. However, file contention for a shared disk decreases performance. In a typical production environment, multiple physical disks on each node are dedicated to the distributed file system, which results in much better performance.

Maximum Disk Allocation: If you are planning to install multiple instances of file system, the number of disks supported on a node can vary based on the number of instances you plan to install. For example, a single node with four instances of the data-fabric FileServer can support up to 360 disks.

Drive Configuration

Do not use RAID or Logical Volume Management with disks that are added to a data-fabric node. While HPE Ezmeral Data Fabric supports these technologies, using them incurs additional setup overhead and can affect your cluster's performance. Due to the possible formatting requirements that are associated with changes to the drive settings, configure the drive settings prior to installing data-fabric.

If you have a RAID controller, disable it, and let the system run in Host Bus Adapter (HBA) mode. For systems that do not support HBA, and have LSI MegaRAID controllers, configure the following drive-group settings for optimal performance:

Property (The actual name depends on the version)	Recommended Setting
RAID Level	RAID0
Stripe Size	>=256K
Cache Policy or I/O Policy	Cached IO or Cached
Read Policy	Always Read Ahead or Read Ahead
Write Policy	Write-Through
Disk Cache Policy or Drive Cache	Disabled

Enabling the Disk Cache policy can improve performance. However, enabling the Disk Cache policy is not recommended because it increases the risk of data loss if the node loses power before the disk cache is committed to disk.

ATTENTION Disable write caching on all data-fabric disks if the disks are not battery backed.

Minimum Disk Space

OS Partition. Provide at least 10 GB of free disk space on the operating system partition. Provide 10 GB of free disk space in the /tmp directory and 128 GB of free disk space in the /opt directory. Services, such as ResourceManager and NodeManager, use the /tmp directory. Files, such as logs and cores, use the /opt directory.

File System. Provide the higher of 8 GB of free disk space or the memory allocated to the data-fabric file system. Note that the disk space should be greater than the memory allocated to the data-fabric file system. If you are using virtual disks, ensure that the virtual disks are thick-provisioned, to avoid the possibility of the virtual disk capacity being greater than actual physical disk capacity. A thin-provisioned virtual disk attempts to write past the end of the physical disk when the physical disk is full. Attempting to write past the end of physical disks could repeat across all thin-provisioned virtual disks at the same time if the thin-provisioned virtual disks are configured in a similar manner.

Swap Space. For production systems, provide at least 4 GB of swap space. If you believe more swap space is needed, consult the swap-space recommendation of your OS vendor. The amount of swap space that a production system needs can vary greatly depending on the application, workload, and amount of RAM in the system. Note that the Installer generates a warning if your swap space is either less than 10% of main memory, or less than 2 GB.

ZooKeeper. On ZooKeeper nodes, dedicate a partition, if practicable, for the /opt/mapr/zkdata directory to avoid other processes filling that partition with writes and to reduce the possibility of errors due to a full /opt/mapr/zkdata directory. This directory is used to store snapshots that are up to 64 MB. Since the four most recent snapshots are retained, reserve at least 500 MB for this partition. Do not share the physical disk where /opt/mapr/zkdata resides with any data-fabric file system data partitions to avoid I/O conflicts that might lead to ZooKeeper service failures.

Virtual Memory (swappiness)

Swappiness is a setting that controls how often the kernel copies the contents of RAM to swap. By setting vm.swappiness to the right value, you can prevent the system from swapping processes too frequently, but still allow for emergency swapping (instead of killing processes). For all Linux distributions, the HPE recommendation is to set vm.swappiness to 1.

To check the current value for vm.swappiness run:

cat /proc/sys/vm/swappiness

To change the value, run:

sudo sysctl vm.swappiness=1

The value of vm.swappiness can revert to a system default setting if you reboot the node. To make this setting permanent, enter vm.swappiness=1 in /etc/sysctl.conf and save it.