| Title: | *OLD* ALL-IN-1 (tm) Support Conference |
| Notice: | Closed - See Note 4331.l to move to IOSG::ALL-IN-1 |
| Moderator: | IOSG::PYE |
| Created: | Thu Jan 30 1992 |
| Last Modified: | Tue Jan 23 1996 |
| Last Successful Update: | Fri Jun 06 1997 |
| Number of topics: | 4343 |
| Total number of notes: | 18308 |
While $ ANAL/RMS/FDL OA$SHARE:OA$DAF_E.DAT, I noticed that the bucket size for this file is 4. On most disks nowadays, most of the time the CLUSTERSIZE is set to 3 by default. In an RMS course that I've attended recently, it was mentioned that the bucket size for a file should always be a multiple of the CLUSTERSIZE for the disk. Since a bucket dictates how much data is read when doing the actual physical I/O, are we not being inneficient here by setting the bucket size to 4? Should it not be set to 6 instead? I can see the case where when adding a DAF record may require the allocation of another data bucket resulting in doing an I/O for 6 blocks for a 4-block bucket. Or have I missed something? Mario
| T.R | Title | User | Personal Name | Date | Lines |
|---|---|---|---|---|---|
| 999.1 | Cluster size not transfer size | IOSG::DAVIS | Mark Davis | Wed Jul 08 1992 10:34 | 21 |
So far as I understand it the disk cluster size has nothing whatsoever
to do with transfer size. In general the disk cluster size should be
ignored as a factor when determining the bucket size of a file.
The disk cluster size does determine the allocation size of the file
which has to be a multiple of the disk cluster size. If you have a
badly fragmented file then buckets may straddle extents leading to
possibly more split I/Os.
You may find it advantageous to increase the bucket size of your SDAF
anyway. If messages have long distribution lists then continuation
SDAF records may be created. You may want to increase the bucket size
to ensure that these records are read in one I/O.
Mark
| |||||
| 999.3 | Instructor's reply. | KAOFS::M_MORIN | Le diable est aux vaches! | Thu Jul 09 1992 16:47 | 21 |
Here is the answer I got back from the instructor in case anyone is curious: Yes, the bucket size should be a multiple of the disk cluster size. However, not because that a disk cluster size is the size of a transfer. For an Index file the size of the transfer is a bucket. The main reason that the cluster size should be a multiple of the disk cluster size is to 1) avoid dead/unusable space in the file, and 2) as an attempt to limit the possibility of split I/Os. Of course all is dependent on an appropriate cluster size haing been selected. An appropriate cluster size is one that is based on the number of sectors in the track, example. Can the number of sectors in a track be evenly divided by the cluster size. If sectors per track is 33, and cluster size is 3. This is a good cluster size, and a good bucket size is also 3. However 6 would not be a good bucket size, because in that sector there would be 5.5 buckets. That last bucket would be a split I/O to process Regards, Mario | |||||