Subject Re: [IB-Architect] UDF replacement: native shared libraries vs. J ava
Author Ann Harrison
At 08:10 AM 4/18/00 +0100, Jason Chapman wrote:
>Why wouldn't it if it is the only significant process running. Moving from
>a 512MB configuration to a 1GB configuration makes a significant improvement
>in performance.

The cache expands up to a preset limit. The limit is expressed as a number
of pages rather than cache bytes. For V6, the limit is 2,000 pages and the
default page size is 4Kb. The superserver allocates one cache per database
and each database gets a cache that size.

The limit on cache size can be overridden in several ways: at the server,
on the connection string (I think)... check the books.

The reason for the limit is that larger caches are not always better.
A mass insert, for example, works better with a small cache since there
is no reason to save filled data pages. At some number around 10,000,
additional cache pages seem to degrade performance. At the moment, I
don't know why. Unlimited cache growth also risks causing the system
to page fault.

The cache is LRU because the last set of tests we did showed that fancier
algorithms increased the cache hit rate, but the cost of maintaining the
information was greater than the saving in I/O time. The cache and the
Unix system cache do, in fact, duplicate each other to some extent.