On Apr 21, 2004, at 20:02, roland wrote:
> please read
> and try to disable CONFIG_BLK_DEV_UBD_SYNC, if it it is enabled for
> your uml.
Oops! Why is it that even when I try searching the mailing list
archives, I can never find what I'm looking for? Thank you for that
reference. It is definitely enabled for my UML (dmesg shows "ubd :
Synchronous mode"). I'll have to try disabling it.
However, I have one question:
[from the above message]:
> Furthermore, I did a few quick tests with this option disabled of
> what happens if the UML is crashed during heavy disk writes (i.e.
> halt command). When using ext3 for the UML's filesystem, I saw no
> filesystem corruption and only very little data loss (which is to be
> expected). Therefore, I see no additional risk to data integrity by
> disabling this option, especially given the performance boost received.
So I have to wonder what will happen if the *host* crashes with this
option. I'm assuming that in this case, the UML will see SEVERE disk
corruption, as the writes, even the ext3 journal writes, will be lost.
In this case, there is probably no point in even using ext3 instead of
ext2, as all of its reliability benefits will be lost. Am I
understanding the choices correctly here? I can either choose to have
my UML have good disk performance, or I can choose it to be reliable,
but not both?
Maybe this could be a future development item for UML? Make a UML-aware
ext3 driver that synchronizes journal writes, but allows the others to
I also found this message:
> I've seen on many other lists reports about IDE disks seeming to be
> faster than SCSI ones. It seems that they report success back to the
> OS as soon as the write is in their buffer, instead of waiting till
> it's in the disk.
I'm guessing that is part of the problem here: The writes are going to
a RAID array, so that is slower than normal already, and on top of that
it actually waits until the writes are complete before reporting back
Additionally, I noticed on the UML diary
(http://user-mode-linux.sourceforge.net/diary.html) that Jeff is
working on AIO, and he hopes that it should improve UML's IO
performance. Can anyone comment on the current status of this? Will it
be possible to run a 2.4 UML on a 2.6 host and take advantage of AIO?
Evan Jones: http://www.eng.uwaterloo.ca/~ejones/
"Computers are useless. They can only give answers" - Pablo Picasso