[Btrfs-devel] Re: Checksum verify failed on btrfs-unstable.

[Evgeniy Polyakov]

Hi Chris.

On Mon, Mar 24, 2008 at 10:01:27AM -0400, Chris Mason (chris.mason at oracle.com) wrote:
> > During stress iozone test I got folowing in the dmesg:
> > btrfs: sdb1 checksum verify failed on 339968 wanted 0 found 4DF48705
> > from_this_trans 0
> 
> These are harmless for now, there is a small window where a buffer can change 
> while it is being written (making the csum invalid).  It is slated to be 
> fixed after the multi device code goes in.

Ah, ok. Thanks for explaination.

> > Performance degraded to hell for this test (8gb file, automatic iozone
> > mode, probably during random read test).
> 
> I'd have to see all the test details to explain performance problems.  I would 
> expect that once you do a series of random writes to the file, sequential 
> reads are going to be painfully slow.  For that kind of database workload, 
> you'll want to mount with -o nodatacow.  Since iozone lays out the file once 
> in the beginning, performance will definitely suffer after it does a random 
> write run.

The same test for 4gb file showed about 2.5 times faster numbers:
4194304      64   75661   73388    74192    74277   47080   69861 
8388608      64   74151   71390    73701    73751   17287   55392
                                                    ^^^^^
I've just checked other filesystems, they have similar drop, that's
because of 4g ram amount, so that most of the data was in the cache for
4g file...

BTRFS just rised an eyebrow because of that messages in dmesg, which
happend to be harmless.

> > Filesystem was made with -l 4k -n 4k options and mounted with
> > noatime,nodatasum,ssd options.
> >
> > Single 266 gb scsi disk.
> > Any additional details?
> 
> The ssd mount option assumes that reads are seek free.  It will be optimal for 
> postmark, which only does random reads and never does sequential reads, but 
> it will be slower for most other benchmarks.

That's for testing only, I will also run all benchmarks without ssd option too.

-- 
	Evgeniy Polyakov