Re: Chunk Writes

> I have been reading the "online Dyn. Server Sys. Adm 1" training > mtrl -by Cindy Lee Dowling,Elizabeth Suto. > > Chapter 10 -270 which discusses Chunk Writes. > > it says " All modified buffers are flushed to disk.The page cleaner > threads read thro' the buffer headers looking for pages associated with > their chunk .At the end of this process , the page cleaner threads sort > all of the pages into Disk Sequence by page numbers and writes this > sequence to disk .This is known as Sorted Write , and it minimises disk > seek time during checkpoint .the writes performed by page cleaners at > Checkpoint time are known as Chunk Writes" > > Now , Does that mean a sequence is written to disk first and then > the pages? Could someone distinguish between Sorted writes and Disk > Writes. In this phase of the checkpoint, only the dirty buffers are written to disk. Other things are written to disk both before and after this step, as is documented in the Dynamic Server Database Server Administrator's Guide Version 7.2 pg. 12-55. When they say the pages are sorted in "Disk Sequence" it is a little misleading. Actually, they are sorted based on page number, and page number is composed of chunk number and page offset within the chunk. Then all of the pages for a chunk are written by a single page cleaner in order within that chunk, which is where the term "Chunk Write" comes from. Chunk write and sorted write are the same thing. Assume, for argument sake, that you have 14 chunks spread across 4 disks and that, based a fairly common rule-of-thumb you have 4 page cleaners. Also assume that each of the 14 chunks has some dirty pages in the buffer cache. The pages for chunk 1 will be assigned to page cleaner 1, chunk 2 to page cleaner 2, and so on until all four cleaners are assigned. The first cleaner to finish its assigned chunk will then be assigned chunk 5, and on and on until all pages have been written to disk. Depending on the distribution of the chunks on the disks, you can have multiple page cleaners writing to the same disk, which does result in contention. This is almost (I'm tempted to leave out the word "almost") unavoidable. Short of actually checking for disk name and then chunk offset within disk, there really is no way to ensure that writing is actually done completely in sequence across the entire disk. Even at that, when you consider logical volumes, disk arrays, disk sections, etc., I doubt that you could ever develop a method that would completely guarantee writes to occur in perfect order across all platforms. Mark Collins mcollins@us.dhl.com The problem lies in how easily and dangerously we forget that manipulating things is not the same as understanding them.