Hi - I just downloaded the sources, as http://sqlite.org/arch.html didn't really say much about the journal or multi-processing.
It appears from looking at pager.c that every process working with a database acquires its own page cache by use of read(). They independently keep their individual cache in sync by playing back the (single) journal. So if you configure a max cache size of M, and you have N persistent processes on a machine, eventually you'll likely get to a N * M total usage of RAM for cache? I can see that this allows for independent readers while one process is writing. But this is at the cost of redundant cache misses, as each reader must do the same reading to refill its cache. I realize that disk pages are likely cached by the OS (in RAM too, after the first processes reads it). So it is likely not to be redundant real disk IO. But it seems like it uses more RAM than it might, if there are multiple persistent processes and a decent sized cache. I realize that alternative architectures add significant complexity: for example, a shared set of pages with shadow versions to maintain transactional isolation, and reference tracking to age away unused pages. (That is, some variant of MVCC.) But is my understanding of the tradeoff correct? -mda --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
