Re: dd from jfs2 to raw very slow..

• Previous message: jean-yves LENHOF: "Re: cygwin and startx"
• In reply to: Tao Chen: "Re: dd from jfs2 to raw very slow.."
• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Date: Sat, 29 Oct 2005 22:13:27 GMT

Tao Chen wrote:
> user wrote:
>
>> Ok.. Does the block size depend on the type of file that you are
>> trying to dd. For example, I'm trying to dd Oracle datafiles, and I
>> was advised to use a 4k block size..
>
>
> No, 128K is chosen because that's the maximum buf size used by LVM
> driver (last time I checked). When you do I/O to raw LVM device,
> 128K is mostly efficient. You probably get better throughput if
> the underneath disks are striped.
>
> It has less effect on files, because filesystem read-ahead code usually
> kicks in when you do sequentail read on a file, no matter JFS or J2.
>
>> Also, does it matter if the raw logical volume is type jfs, as opposed
>> to jfs2..
>
>
> No, LV 'type' doesn't affect anything, AFAIK.

Actually is does matter. raw logical volumes do not use the jfslog,
while jfs and jfs2 filesystems do. Also, you have other mount options
with jfs and jfs2 (dio,cio,nomind can also be set for jfs), that can
affect the performance. I/O to raw lvs also bypasses vmm and is a direct
write to a disk. If you are using jfs, you cannot use dio with a large
file enabled filesystem as it actually slows down the write, you would
either have to use jfs2 or a non-LFE filesystem to get the benefits of
dio. cio, can only be used with jfs2, last I checked.
>
>> Will that (having the raw logical volume having a different type (jfs)
>> effect whether or not the new file can be read or not ?
>
>
> Supposedly you will reconfigure Oracle to use the copy on raw LV?
> I don't know if Oracle cares about LV 'type' at all.
>
> Tao
>

• Previous message: jean-yves LENHOF: "Re: cygwin and startx"
• In reply to: Tao Chen: "Re: dd from jfs2 to raw very slow.."
• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]