On 23 Apr 2002 15:10:03 GMT, chao_ping <chao_ping_at_163.com> wrote:
>Since some unix filesystem like sun ufs and veritas vxfs now support
>direct I/O, from steve adams's site, it seems that using direct I/O is
>better than use buffered I/O. For example, i have a system like: 8CPU/8G
>mem/2 T3 Raid5/Veritas VXFS. This is an OLTP system, most wait event is
>db file sequential read. Which option shall i use? use direct I/O, or
>just general buffered I/O?Any one have some experience? And if using
>Quick I/O, since it act like raw device,How much memory shall i leave
>for the filesystem buffer ?
db file sequential read is either random single block reads or
transfers from the TEMP tablespace during a sort. The TEMP tablespace
I/O might benefit from buffering, but odds are the random single block
reads won't see any benefit from system buffers assuming that your
raid hardware has a write cache (writes to RAID 5 require at least 4
i/o operations). If you do not have a write cache, you need to leave
the system buffer cache on. Otherwise, switch to direct i/o, if you
notice a slow down in sorts that have to go to disk, turn the buffers
back on for the temp tablespace (if it is on a seperate file system)
or consider increasing your sort area size (which implies rebuilding
your TEMP tablespace to keep the extent size appropriate).
Received on Tue Apr 23 2002 - 12:09:08 CDT
