Am 19.04.2010 23:09, schrieb Andrew Morton:
>
> (switched to email. Please respond via emailed reply-to-all, not
> via the bugzilla web interface).
ok.
> Sigh. This shouldn't happen.
>
> I'm going to go ahead and assume that some earlier kernels didn't
> do this :(
No, I don't _think_ so. Problem is, I only noticed this when zeroing out
the device. Haven't done this recently with earlier kernels. Besides,
as reported, not all write operations are slow.
> Is the writeout to /dev/sde1 slow right from the start, or does it
> start out fast and later slow down?
It seems to me that write speed does not change very much, but the
actual speed varies. One time it's just 10 MiB/s, another time about 30.
But for a couple of tests writes, speed is pretty stable.
Also, only writing directly to the device seems to be slow.
Perhaps that's why is hard to notice.
e.g.: when /dev/sde1 is mounted as ext3 fs:
Copying a 5 GiB file takes about 50s,
mbuffer </dev/zero >/mnt1/foo1 writes with >100 MiB/s,
dd if=/dev/zero of=/mnt1/foo2 manages >75 MiB/s.
After unmounting and writing directly from /dev/zero to /dev/sde1:
mbuffer writes about 40 MiB/s and dd less than 30.
But then badblocks -svw -t 0x00 /dev/sde1 writes about 120 MiB/s.
Btw, dd if=/dev/zero of=/dev/null runs with >500 MiB/s.
Needless to say, I'm confused by the numbers...
> `dd' isn't very efficient without the `bs' option - it reads and
> writes in 512-byte chunks. But that shouldn't be causing these
> problems.
Tried mbuffer instead. Shows similar results, only a bit faster.
Still, even if dd is inefficient, it is way slower than expected.
If you want me to test anything, please let me know.
Walter
PS: The short tests above triggered no "page allocation failures"
(vm.vfs_cache_pressure = 1000).
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>