Operating System Configuration
We recommend to not use BTRFS on linux, it's known to not work well in conjunction with ArangoDB. We experienced that arangodb facing latency issues on accessing its database files on BTRFS partitions. In conjunction with BTRFS and AUFS we also saw data loss on restart.
Virtual Memory Page Sizes
By default, ArangoDB uses Jemalloc as the memory allocator. Jemalloc does a good job of reducing virtual memory fragmentation, especially for long-running processes. Unfortunately, some OS configurations can interfere with Jemalloc's ability to function properly. Specifically, Linux's "transparent hugepages", Windows' "large pages" and other similar features sometimes prevent Jemalloc from returning unused memory to the operating system and result in unnecessarily high memory use. Therefore, we recommend disabling these features when using Jemalloc with ArangoDB. Please consult your operating system's documentation for how to do this.
sudo bash -c "echo madvise >/sys/kernel/mm/transparent_hugepage/enabled" sudo bash -c "echo madvise >/sys/kernel/mm/transparent_hugepage/defrag"
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 from killing ArangoDB too eagerly on Linux.
The recommended kernel setting for
overcommit_memory for both MMFiles and
RocksDB storage engine is 0 or 1. The kernel default is 0.
You can set it as follows before executing
sudo bash -c "echo 0 >/proc/sys/vm/overcommit_memory"
When this flag is 0, the kernel attempts to estimate the amount of free memory left when userspace requests more memory.
When this flag is 1, the kernel pretends there is always enough memory 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.
sudo bash -c "echo 0 >/proc/sys/vm/zone_reclaim_mode"
This is value ORed together of
- 1 = Zone reclaim on
- 2 = Zone reclaim writes dirty pages out
- 4 = Zone reclaim swaps pages
Multi-processor systems often have non-uniform Access Memory (NUMA). ArangoDB should be started with interleave on such system. This can be achieved using
numactl --interleave=all arangod ...
Max Memory Mappings
Linux kernels by default restrict the maximum number of memory mappings of a single process to about 64K mappings. While this value is sufficient for most workloads, it may be too low for a process that has lots of parallel threads that 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 of mappings is reached, calls to mmap will fail, so the process will think no more memory is available although there may be plenty of RAM left.
To avoid this scenario, it is recommended to raise the default value for the maximum number of memory mappings to a sufficiently high value. As a rule of thumb, 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 higher value for the number of memory mappings.
To set the value once, use the following command before starting arangod:
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 is looking at.
It is recommended to set the environment variable
GLIBCXX_FORCE_NEW to 1 on
systems that use glibc++ in order to disable the memory pooling built into
glibc++. That memory pooling is unnecessary because Jemalloc will already do
While it is possible to compile ArangoDB on 32bit system, this is not a recommended environment. 64bit systems can address a significantly bigger memory region.