Zhihui Zhang wrote: > My feeling is that if we allocate ALL the data blocks of a big file > contiguously, this will lead to "too much localization" as described in > the paper (or the book). However, this may be good for this big file if > the system buffering capability and hardware allow it (at the cost of > other files?)
Maybe this is something we could get if XFS is ported: XFS's guaranteed rate I/O (partly) works by putting guaranteed-rate files on distinct positions on the disk, or different "subvolumes" in the case of GRIO on XLV logical volumes. So when preparing a filesystem you could build a logical volume out of twenty 9 Gbyte disks plus another five 9 Gbyte disks for guaranteed-rate files. [ in practice you'd probably be building such a filesystem for a specific application, though, so you'd probably really use 25 9 Gbyte disks for GRIO :-) ] You decide which subvolume a file is allocated to immediately after file creation: There's an ioctl() which can be used before the first write to a new file which sets the "please make me fast" flag. One thing that helps to make this possible is an I/O scheduler which supports prioritization. Hmm... - mark ---- Mark Newton Email: new...@internode.com.au (W) Network Engineer Email: new...@atdot.dotat.org (H) Internode Systems Pty Ltd Desk: +61-8-82232999 "Network Man" - Anagram of "Mark Newton" Mobile: +61-416-202-223 To Unsubscribe: send mail to majord...@freebsd.org with "unsubscribe freebsd-hackers" in the body of the message