Hi:
I too will adopt the radical position that bufpages should be non-zero and thus a static buffer cache. I will say that your value of bufpages=9000 (36MB) is probably too low but since you failed to mention anything about your resources or applications; it's impossible to say with any degree of confidence. In most cases, I set nbuf to 0; the default header configuration works well.
In the vast majority of cases setting bufpages above about 80000 (320 MB or so) results in only very marginal increases in buffer cache hit rates. A few boxes (11x) will benefit from buffer caches in the 800-900 MB range but that is the exception rather than the rule.
One of the main reasons that I prefer static buffer caches is that it is much easier to see the effect of other tunables when this value is held constant. I can also say the the very best system tuner I know will tell you to use static buffer cache even though in his HP politically-correct paper, he suggests that you use dynamic buffer cache. When he is explaining this, he makes rapid up-and-down motions with his hand and explains the kernel overhead associated with it.
Of course, the best answer is to try both and measure performance. If you choose dynamic buffer cache, I would set min_dbc_pct to about 5 and max_dbc_pct to between 10 and 20.
If this is a pure database server doing lots of raw/io (or the Online JFS equivalent), then the buffer cache can be reduced quite low. This might explain your current setting although I generally set bufpages at least a bit higher.
Food for thought, Clay
If it ain't broke, I can fix that.
