Knowledge Base

How to estimate initial memory configuration

The initial and eventual memory configuration parameters can be a moving target, based on how your store changes in size and how your workload increases or changes over time.

This guidance is meant for the initial configuration.

In order to decide on an appropriate configuration, you need the following information:

  • Amount of physical memory on the machine hosting Neo4j.

  • Estimates of the following:

    • Number of nodes.

    • Number of relationships.

    • Average number of properties per node and per relationship.

A fairly high-level rule of thumb is Total Physical Memory = Heap + Page Cache + OS Memory.

Usually reserving 1-2GB for the OS is sufficient. The heap and page cache are detailed below.

First, we need to decide on a good heap size.

A heap should not be overly large, as that can cause much longer Stop-the-World pauses when a full garbage collection (GC) cycle is needed. It should also be big enough to allow enough memory for your workload. On systems with a large amount of physical memory (>56GB), keeping the heap to 16GB and under typically works well.

Second, consider the page cache.

This is where the store files will be mapped in main memory for quicker access. The default page-cache size is 50% of the available memory. A good rule of thumb is Store size + expected growth + 10%. So, for a store that is 5GB in size, and you expect that to double in size in the next year, you would ideally allocate 5GB + 5GB + 1GB = 11GB.

This last section is no longer relevant for Neo4j versions 2.3 and later.

Lastly, let’s look at the object cache options.

The object cache is where nodes, relationships, and other objects are mapped to main memory. By default on Neo4j 2.2.x, this is set to High Performance Cache (hpc). On small stores (~10GB or smaller), this performs well. On larger stores, you will probably see better performance turning off the cache (set cache_type=none).

If you are using the object cache and need to tune it further, consider starting with the cache.memory_ratio option. This is on-heap, so it is the percentage of the heap to use for object cache. The default is 50%, but you can increase this a bit (to as high as 65-70%), especially if the JVM is not using all of its heap consistently.