A few filesystem benchmarks w/ReiserFS4 vs Other Filesystems

A few filesystem benchmarks w/ReiserFS4 vs Other Filesystems

[Justin Piszcz]

Are balanced b-trees better for removing many files over dancing onces?
See rm -rf benchmark.

# -------------------------------------------------------------------- #
Filesystems to test:
# -------------------------------------------------------------------- #
Each file system was created with fdisk to be +1024MB.
# -------------------------------------------------------------------- #
With default initilization commands: mkfs.fs /dev/hdb[1-7]
# -------------------------------------------------------------------- #
Filesystem            Size  Used Avail Use% Mounted on
/dev/hdb1             962M   20K  913M   1% /fs/ext2
/dev/hdb2             962M   17M  897M   2% /fs/ext3
/dev/hdb3             977M   33M  945M   4% /fs/reiser3
/dev/hdb5             929M  144K  929M   1% /fs/reiser4
/dev/hdb6             973M  256K  973M   1% /fs/jfs
/dev/hdb7             973M  144K  972M   1% /fs/xfs
# -------------------------------------------------------------------- #
Untar the Linux 2.6.8.1 tarball on each file system.
# -------------------------------------------------------------------- #
ext2 | 46.88 sec @ 10% cpu
ext3 | 44.44 sec @ 12% cpu
  jfs | 57.36 sec @ 15% cpu
  rs3 | 37.03 sec @ 23% cpu
  rs4 | 27.42 sec @ 42% cpu
  xfs | 49.74 sec @ 17% cpu
# -------------------------------------------------------------------- #
Execute rm -rf linux-2.6.8.1 on each file system.
# -------------------------------------------------------------------- #
ext2 | 10.26 sec @ 22% cpu
ext3 | 10.02 sec @ 25% cpu
  jfs | 26.67 sec @ 27% cpu
  rs3 | 03.22 sec @ 74% cpu
  rs4 | 25.58 sec @ 50% cpu <- What happened to reiserfs4 here?
  xfs | 12.51 sec @ 47% cpu
# -------------------------------------------------------------------- #
Create a 500MB file with dd to each filesystem with 1MB blocks.
# -------------------------------------------------------------------- #
ext2 | 15.72 sec @ 26% cpu
ext3 | 17.04 sec @ 31% cpu
  jfs | 29.57 sec @ 25% cpu
  rs3 | 15.21 sec @ 27% cpu
  rs4 | 23.96 sec @ 23% cpu <- What happened to reiserfs4 here?
  xfs | 19.07 sec @ 29% cpu

Re: A few filesystem benchmarks w/ReiserFS4 vs Other Filesystems

[Bartlomiej Zolnierkiewicz]

Hi,

On Friday 27 August 2004 16:39, Justin Piszcz wrote:
> Are balanced b-trees better for removing many files over dancing onces?
> See rm -rf benchmark.
>
> # -------------------------------------------------------------------- #
> Filesystems to test:
> # -------------------------------------------------------------------- #
> Each file system was created with fdisk to be +1024MB.
> # -------------------------------------------------------------------- #
> With default initilization commands: mkfs.fs /dev/hdb[1-7]
> # -------------------------------------------------------------------- #
> Filesystem            Size  Used Avail Use% Mounted on
> /dev/hdb1             962M   20K  913M   1% /fs/ext2
> /dev/hdb2             962M   17M  897M   2% /fs/ext3
> /dev/hdb3             977M   33M  945M   4% /fs/reiser3
> /dev/hdb5             929M  144K  929M   1% /fs/reiser4
> /dev/hdb6             973M  256K  973M   1% /fs/jfs
> /dev/hdb7             973M  144K  972M   1% /fs/xfs

Sorry to say this but your testing procedure is flakey because outer disk 
tracks are much faster so you should repeat all tests using different 
filesystems on i.e. only /dev/hdb1.  The other thing is to get caching out of 
picture - you should do reboots between tests - create fs, reboot, test, 
create fs... cycle.

PS my mail account is temporary defunct so please use Reply-To:

> # -------------------------------------------------------------------- #
> Untar the Linux 2.6.8.1 tarball on each file system.
> # -------------------------------------------------------------------- #
> ext2 | 46.88 sec @ 10% cpu
> ext3 | 44.44 sec @ 12% cpu
>   jfs | 57.36 sec @ 15% cpu
>   rs3 | 37.03 sec @ 23% cpu
>   rs4 | 27.42 sec @ 42% cpu
>   xfs | 49.74 sec @ 17% cpu
> # -------------------------------------------------------------------- #
> Execute rm -rf linux-2.6.8.1 on each file system.
> # -------------------------------------------------------------------- #
> ext2 | 10.26 sec @ 22% cpu
> ext3 | 10.02 sec @ 25% cpu
>   jfs | 26.67 sec @ 27% cpu
>   rs3 | 03.22 sec @ 74% cpu
>   rs4 | 25.58 sec @ 50% cpu <- What happened to reiserfs4 here?
>   xfs | 12.51 sec @ 47% cpu
> # -------------------------------------------------------------------- #
> Create a 500MB file with dd to each filesystem with 1MB blocks.
> # -------------------------------------------------------------------- #
> ext2 | 15.72 sec @ 26% cpu
> ext3 | 17.04 sec @ 31% cpu
>   jfs | 29.57 sec @ 25% cpu
>   rs3 | 15.21 sec @ 27% cpu
>   rs4 | 23.96 sec @ 23% cpu <- What happened to reiserfs4 here?
>   xfs | 19.07 sec @ 29% cpu

Re: A few filesystem benchmarks w/ReiserFS4 vs Other Filesystems

[Alexander Lyamin]

Fri, Aug 27, 2004 at 05:45:41PM +0200, Bartlomiej Zolnierkiewicz wrote:
> 
> Hi,
> > Execute rm -rf linux-2.6.8.1 on each file system.
> > # -------------------------------------------------------------------- #
> > ext2 | 10.26 sec @ 22% cpu
> > ext3 | 10.02 sec @ 25% cpu
> >   jfs | 26.67 sec @ 27% cpu
> >   rs3 | 03.22 sec @ 74% cpu
> >   rs4 | 25.58 sec @ 50% cpu <- What happened to reiserfs4 here?
> >   xfs | 12.51 sec @ 47% cpu
> > # -------------------------------------------------------------------- #
> > Create a 500MB file with dd to each filesystem with 1MB blocks.
> > # -------------------------------------------------------------------- #
> > ext2 | 15.72 sec @ 26% cpu
> > ext3 | 17.04 sec @ 31% cpu
> >   jfs | 29.57 sec @ 25% cpu
> >   rs3 | 15.21 sec @ 27% cpu
> >   rs4 | 23.96 sec @ 23% cpu <- What happened to reiserfs4 here?
> >   xfs | 19.07 sec @ 29% cpu

Your answers somewhere in HCH's "silent semantics" thread.

Basically reiserfs team aware that they do suck at file DELETES
and OVERWRITES.  There seem to be a way to rectify this perfomance
issues in future (dynamic repacker?). Altough i was somewhat surprised
with this  dd file benchmark... probably Alexander Zarochentsev knows
the answer.
-- 
"the liberation loophole will make it clear.."
lex lyamin

Re: A few filesystem benchmarks w/ReiserFS4 vs Other Filesystems

[Justin Piszcz]

Thanks for the feedback.

On Fri, 3 Sep 2004, Alexander Lyamin wrote:

> Fri, Aug 27, 2004 at 05:45:41PM +0200, Bartlomiej Zolnierkiewicz wrote:
>>
>> Hi,
>>> Execute rm -rf linux-2.6.8.1 on each file system.
>>> # -------------------------------------------------------------------- #
>>> ext2 | 10.26 sec @ 22% cpu
>>> ext3 | 10.02 sec @ 25% cpu
>>>   jfs | 26.67 sec @ 27% cpu
>>>   rs3 | 03.22 sec @ 74% cpu
>>>   rs4 | 25.58 sec @ 50% cpu <- What happened to reiserfs4 here?
>>>   xfs | 12.51 sec @ 47% cpu
>>> # -------------------------------------------------------------------- #
>>> Create a 500MB file with dd to each filesystem with 1MB blocks.
>>> # -------------------------------------------------------------------- #
>>> ext2 | 15.72 sec @ 26% cpu
>>> ext3 | 17.04 sec @ 31% cpu
>>>   jfs | 29.57 sec @ 25% cpu
>>>   rs3 | 15.21 sec @ 27% cpu
>>>   rs4 | 23.96 sec @ 23% cpu <- What happened to reiserfs4 here?
>>>   xfs | 19.07 sec @ 29% cpu
>
> Your answers somewhere in HCH's "silent semantics" thread.
>
> Basically reiserfs team aware that they do suck at file DELETES
> and OVERWRITES.  There seem to be a way to rectify this perfomance
> issues in future (dynamic repacker?). Altough i was somewhat surprised
> with this  dd file benchmark... probably Alexander Zarochentsev knows
> the answer.
> -- 
> "the liberation loophole will make it clear.."
> lex lyamin
>

Re: A few filesystem benchmarks w/ReiserFS4 vs Other Filesystems

[Hans Reiser]

Alexander Lyamin wrote:

>Fri, Aug 27, 2004 at 05:45:41PM +0200, Bartlomiej Zolnierkiewicz wrote:
>  
>
>>Hi,
>>    
>>
>>>Execute rm -rf linux-2.6.8.1 on each file system.
>>># -------------------------------------------------------------------- #
>>>ext2 | 10.26 sec @ 22% cpu
>>>ext3 | 10.02 sec @ 25% cpu
>>>  jfs | 26.67 sec @ 27% cpu
>>>  rs3 | 03.22 sec @ 74% cpu
>>>  rs4 | 25.58 sec @ 50% cpu <- What happened to reiserfs4 here?
>>>  xfs | 12.51 sec @ 47% cpu
>>># -------------------------------------------------------------------- #
>>>Create a 500MB file with dd to each filesystem with 1MB blocks.
>>># -------------------------------------------------------------------- #
>>>ext2 | 15.72 sec @ 26% cpu
>>>ext3 | 17.04 sec @ 31% cpu
>>>  jfs | 29.57 sec @ 25% cpu
>>>  rs3 | 15.21 sec @ 27% cpu
>>>  rs4 | 23.96 sec @ 23% cpu <- What happened to reiserfs4 here?
>>>      
>>>
Do a dd of a 50GB file, I expect a completely different result.  
Basically, this is an artifact of reiser4 choosing to flush the whole 
file once it starts to flush.

>>>  xfs | 19.07 sec @ 29% cpu
>>>      
>>>
>
>Your answers somewhere in HCH's "silent semantics" thread.
>
>Basically reiserfs team aware that they do suck at file DELETES
>and OVERWRITES.  There seem to be a way to rectify this perfomance
>issues in future (dynamic repacker?). Altough i was somewhat surprised
>with this  dd file benchmark... probably Alexander Zarochentsev knows
>the answer.
>  
>