IBM 000-8697 User Manual

IBM Informix OnLine Database Server Administrator’s Guide

Data is always written to the primary chunk first. If a mirror chunk is
associated with the primary chunk, the write is repeated on the mirror chunk.
The write to the mirror chunk is also included in these counts.

Refer to

 for a discussion of tuning OnLine performance by

monitoring write-type statistics.

Refer to

for information about monitoring write types (and buffer

flushing) using tbstat -F.

Any OnLine process that is writing more than one page to the same disk sorts
the pages into disk sequence and writes the pages in disk-sequence order.
(The disk-sequence information is contained in the physical address of the
page, which is contained in the page header.) This technique is referred to as
a sorted write.

Sorted writes are more efficient than either idle writes or foreground writes
because they minimize head movement (disk seek time) on the disk. In
addition, a sorted write enables the page cleaners to use the big buffers
during the write, if possible. (Refer to

for more information about

big buffers.)

Chunk writes, which occur during checkpoints, are performed as sorted
writes. (Chunk writes are the most efficient writes available to OnLine. Refer
to

.)

Writes that are initiated by the page cleaners are called idle writes. The page-
cleaner daemon wakes periodically and searches through the

MLRU

 queues

to determine if the number of modified buffers is equal to or greater than the
value represented by

LRU_MAX_DIRTY

.

If a page cleaner determines that the buffer pool should be flushed, it marks
a page for flushing, flushes the page (after first checking to determine if the
physical log buffer must be flushed first), and then rechecks the percentage.
This process repeats until the number of modified buffers is less than the
value represented by

LRU_MIN_DIRTY

.