On Tue, 2003-08-19 at 03:12, Aaron Hope wrote: > I think that he's looking for the same raw i/o that oracle likes so > much: > http://www.tldp.org/HOWTO/SCSI-2.4-HOWTO/rawdev.html
Yep. I've also seen this reference. > > But while you don't have to be writing a filesystem to want this, it's > still targeted for a pretty select audience. Have you considered a > simple O_DIRECT instead? I'm running 2.6.0-test3 so I think O_DIRECT may be supported. I'm under the assumption that RAW i/o came before memory mapped i/o. Seems RAW uses user space memory while memory mapped uses kernel space memory? Is that right? I also found a paper that showed that RAW was actually slower than EXT3 for a database system. So advantages of RAW: * data is definitely written to disk when write/pwrite returns (what about the disk's cache?) * no extra buffer * no file system overhead Disadvantages: * must read/write in proper block size bytes * no file system - it is just a stream of bytes And to use a raw device: * use raw to associate a block device with a raw device * open() the raw device * read/write to it - use seek to move around Please correct me if I have the wrong idea. I do have Stevens' books and consider them to be my bibles. :-) _______________________________________________ gnhlug-discuss mailing list [EMAIL PROTECTED] http://mail.gnhlug.org/mailman/listinfo/gnhlug-discuss
