A person to ask would be Ingo Macherius. However in 2009 we defaulted the cache size to ~21mb (2^20 * 20 bytes) based on data done at that time. A cache entry is 20 bytes in size.
Many things have changed since then (Including a major caching bug that I fixed later in 2009). However for sake of brevity here was Ingo found. He took 5 different 7pt matches on Fibs and analyzed each 4 times (2ply/no pruning/cubeful) - so a total of 20 matches analyzed. He did that for each of the cache sizes and also for differing number of threads on his 4 core system. He used the command line version on Linux and cleared the cache between each analysis.
In the diagram the cache size can be converted to bytes with 2^(value of x scale)*20. Overall the best overall performance was 21mb (2^20 * 20). (These are the values you see on the Cache setting slider in GNUBG).
Whether these cache settings hold up today - good question. I'll probably create some scripts and burn some CPU time to find out. As well I will probably benchmark some Rollouts.
Some extra info can be found from this mailing list thread: http://www.mail-archive.com/bug-gnubg@gnu.org/msg04966.html