Selecting a cache size
The size of the cache used for the underlying database can be specified by calling the DB->set_cachesize method. Choosing a cache size is, unfortunately, an art. Your cache must be at least large enough for your working set plus some overlap for unexpected situations.
When using the Btree access method, you must have a cache big enough for the minimum working set for a single access. This will include a root page, one or more internal pages (depending on the depth of your tree), and a leaf page. If your cache is any smaller than that, each new page will force out the least-recently-used page, and Berkeley DB will re-read the root page of the tree anew on each database request.
If your keys are of moderate size (a few tens of bytes) and your pages are on the order of 4K to 8K, most Btree applications will be only three levels. For example, using 20 byte keys with 20 bytes of data associated with each key, a 8KB page can hold roughly 400 keys (or 200 key/data pairs), so a fully populated three-level Btree will hold 32 million key/data pairs, and a tree with only a 50% page-fill factor will still hold 16 million key/data pairs. We rarely expect trees to exceed five levels, although Berkeley DB will support trees up to 255 levels.
The rule-of-thumb is that cache is good, and more cache is better. Generally, applications benefit from increasing the cache size up to a point, at which the performance will stop improving as the cache size increases. When this point is reached, one of two things have happened: either the cache is large enough that the application is almost never having to retrieve information from disk, or, your application is doing truly random accesses, and therefore increasing size of the cache doesn't significantly increase the odds of finding the next requested information in the cache. The latter is fairly rare -- almost all applications show some form of locality of reference.
That said, it is important not to increase your cache size beyond the capabilities of your system, as that will result in reduced performance. Under many operating systems, tying down enough virtual memory will cause your memory and potentially your program to be swapped. This is especially likely on systems without unified OS buffer caches and virtual memory spaces, as the buffer cache was allocated at boot time and so cannot be adjusted based on application requests for large amounts of virtual memory.
For example, even if accesses are truly random within a Btree, your access pattern will favor internal pages to leaf pages, so your cache should be large enough to hold all internal pages. In the steady state, this requires at most one I/O per operation to retrieve the appropriate leaf page.
You can use the db_stat utility to monitor the effectiveness of your cache. The following output is excerpted from the output of that utility's -m option:
prompt: db_stat -m 131072 Cache size (128K). 4273 Requested pages found in the cache (97%). 134 Requested pages not found in the cache. 18 Pages created in the cache. 116 Pages read into the cache. 93 Pages written from the cache to the backing file. 5 Clean pages forced from the cache. 13 Dirty pages forced from the cache. 0 Dirty buffers written by trickle-sync thread. 130 Current clean buffer count. 4 Current dirty buffer count.
The statistics for this cache say that there have been 4,273 requests of the cache, and only 116 of those requests required an I/O from disk. This means that the cache is working well, yielding a 97% cache hit rate. The db_stat utility will present these statistics both for the cache as a whole and for each file within the cache separately.