Starting with version 1.0 Galera stores writesets in a special cache. It's purpose is to improve control of Galera memory usage and offload writeset storage to disk (see also this faq). Galera cache has 3 types of stores:
A permanent in-memory store, where writesets are allocated by a default OS
memory allocator. It can be useful for systems with spare RAM. It has a hard size limit. By default it is disabled (size set to 0).
A permanent ring-buffer file which is preallocated on disk during cache initialization. It is intended as the main writeset store. By default its size is 128Mb.
An on-demand page store, which allocates memory-mapped page files during runtime as necessary. Default page size is 128Mb, but it can be bigger if it needs to store a bigger writeset. The size of page store is limited by the free disk space. By default page files are deleted when not in use, but a limit can be set on the total size of the page files to keep. When all other stores are disabled, at least one page file is always present on disk.
Allocation algorithm is as follows: all stores are tried in the above order. If a store does not have enough space to allocate the writeset, then the next store is tried. Page store should always succeed unless the writeset is bigger than the available disk space.
By default Galera cache allocates files in the working directory of the process, but a dedicated location can be specified. See Galera parameters list.
Since all cache files are memory-mapped, the process may appear to use more memory than it actually does.