Linux Operating System Configuration

The most important suggestions listed in this section can beeasily applied by making use of a script. Please refer to the pageLinux OS Tuning Script Examples forready-to-use examples.

File Systems

We recommend to not use BTRFS on linux, as it is known to not workwell in conjunction with ArangoDB. We experienced that ArangoDBfacing latency issues on accessing its database files on BTRFSpartitions. In conjunction with BTRFS and AUFS we also saw data losson restart.

Virtual Memory Page Sizes

By default, ArangoDB uses Jemalloc as the memory allocator. Jemalloc does a goodjob of reducing virtual memory fragmentation, especially for long-runningprocesses. Unfortunately, some OS configurations can interfere with Jemalloc’sability to function properly. Specifically, Linux’s “transparent hugepages”,Windows’ “large pages” and other similar features sometimes prevent Jemallocfrom returning unused memory to the operating system and result in unnecessarilyhigh memory use. Therefore, we recommend disabling these features when usingJemalloc with ArangoDB. Please consult your operating system’s documentation forhow to do this.

Execute:

  1. sudo bash -c "echo madvise >/sys/kernel/mm/transparent_hugepage/enabled"
  2. sudo bash -c "echo madvise >/sys/kernel/mm/transparent_hugepage/defrag"

before executing arangod.

Swap Space

It is recommended to assign swap space for a server that is running arangod.Configuring swap space can prevent the operating system’s OOM killer fromkilling ArangoDB too eagerly on Linux.

Over-Commit Memory

The recommended kernel setting for overcommit_memory for both MMFiles andRocksDB storage engine is 0 or 1. The kernel default is 0.

You can set it as follows before executing arangod:

  1. sudo bash -c "echo 0 >/proc/sys/vm/overcommit_memory"

From www.kernel.org:

  • When this flag is 0, the kernel attempts to estimate the amountof free memory left when userspace requests more memory.

  • When this flag is 1, the kernel pretends there is always enoughmemory until it actually runs out.

  • When this flag is 2, the kernel uses a “never overcommit”policy that attempts to prevent any overcommit of memory.

Zone Reclaim

Execute

  1. sudo bash -c "echo 0 >/proc/sys/vm/zone_reclaim_mode"

before executing arangod.

From www.kernel.org:

This is value ORed together of

  • 1 = Zone reclaim on
  • 2 = Zone reclaim writes dirty pages out
  • 4 = Zone reclaim swaps pages

NUMA

Multi-processor systems often have non-uniform Access Memory (NUMA). ArangoDBshould be started with interleave on such system. This can be achieved using

  1. numactl --interleave=all arangod ...

Max Memory Mappings

Linux kernels by default restrict the maximum number of memory mappings of asingle process to about 64K mappings. While this value is sufficient for mostworkloads, it may be too low for a process that has lots of parallel threadsthat all require their own memory mappings. In this case all the threads’ memory mappings will be accounted to the single arangod process, and the maximum number of 64K mappings may be reached. When the maximum number ofmappings is reached, calls to mmap will fail, so the process will think nomore memory is available although there may be plenty of RAM left.

To avoid this scenario, it is recommended to raise the default value for themaximum number of memory mappings to a sufficiently high value. As a rule ofthumb, one could use 8 times the number of available cores times 8,000.

For a 32 core server, a good rule-of-thumb value thus would be 2,048,000 (32 8 8000). For certain workloads, it may be sensible to use even a highervalue for the number of memory mappings.

To set the value once, use the following command before starting arangod:

  1. sudo bash -c "sysctl -w 'vm.max_map_count=2048000'"

To make the settings durable, it will be necessary to store the adjusted settings in /etc/sysctl.conf or other places that the operating system islooking at.

Environment Variables

It is recommended to set the environment variable GLIBCXX_FORCE_NEW to 1 onsystems that use glibc++ in order to disable the memory pooling built intoglibc++. That memory pooling is unnecessary because Jemalloc will already domemory pooling.

Execute

  1. export GLIBCXX_FORCE_NEW=1

before starting arangod.

32bit

While it is possible to compile ArangoDB on 32bit system, this is not arecommended environment. 64bit systems can address a significantly biggermemory region.