Eric Bollengier wrote:
> Le Friday 07 November 2008 14:36:46 Kern Sibbald, vous avez écrit :
>> On Friday 07 November 2008 13:19:45 Ralf Gross wrote:
>>> Kern Sibbald schrieb:
>>>> On Thursday 06 November 2008 22:47:41 Ralf Gross wrote:
>>>>> Alex Chekholko schrieb:
>>>>>> On Wed, 5 Nov 2008 16:12:51 +0100
>>>>>>
>>>>>> Kern Sibbald <kern AT sibbald DOT com> wrote:
>>>>>>> For writing to tape (providing it is LTO-n) I strongly recommend
>>>>>>> a block size not to exceed 256K.
>>>>>> Hi Kern,
>>>>>>
>>>>>> Why do you say that? Is this thread relevant?:
>>>>>> http://www.mail-archive.com/bacula-devel AT lists.sourceforge DOT
>>>>>> net/msg0
>>>>>> 12 46.h tml
>>>>>>
>>>>>> Also, I would like to corroborate the OP's experiences; I had an
>>>>>> almost identical thread about small block size and slow write
>>>>>> speed: http://www.nabble.com/LTO-4-performance--td17407840.html
>>>>>>
>>>>>> In fact, I was unable to get higher block sizes working at all with
>>>>>> btape:
>>>>>> http://www.adsm.org/lists/html/Bacula-users/2008-05/msg00504.html
>>>>>>
>>>>>> So I am still stuck at ~22MB/s writing to LTO-4 with the default
>>>>>> block size.
>>>>> I don't think that the blocksize is the problem. I did some tests but
>>>>> couldn't get higher results with larger blocksizes. I get 75-85 MB/s
>>>>> with the default bs and no additional tuning.
>>>> That is probably correct, but most likely only because you have a
>>>> bottleneck elsewhere -- probably in one of the points I mentioned. The
>>>> speed is always capped by the slowest component. Once you remove the
>>>> other bottlenecks on your system, the blocksize will very likely become
>>>> the bottleneck and then you can measure the difference.
>>> I didn't want to compain, just show the org. poster that his 22 MB/s
>>> are likely not a bs issue.
>>>
>>> That being said, I started a thread on the user list a while ago where
>>> I aked what throughput people are getting when writing to tape. Nobody
>>> involved in this thread got higher numbers than 80-85 MB/s for a
>>> single job.
>> That is probably reasonable for one job, but if you are writing to an
>> LTO-2,3, or 4, we know that with multiple simultaneous jobs it is possible
>> to get write speeds of 150 MB/sec.
>>
>> Kern
>
> This is with a good hardware compression rate, but it's a very good test,
> IMHO, more than using random data to get only 80MB/sec. If you able to write
> at 150MB/s, i'm pretty sure that you will be able to write at 80MB/s... Even
> if your source file is made with a dd if=/dev/zero, your harddisk, your
> network, your SCSI/SAS or whatever controler have to handle it like real
> data.
>
>
I've done extensive throughput experiments on my Dell LTO4 and have
never managed to get the advertised streaming-rate of 120MBytes/sec.
Even writing a stream ov zeros to the drive with hardware-compression
turned on only gets me 100MBytes/sec.
All zeros without compression:
[root]# mt -f /dev/st0 compression 0
[root]# dd if=/dev/zero of=/dev/nst0 bs=1000000 count=10000
10000000000 bytes (10 GB) copied, 107.498 seconds, 93.0 MB/s
All zeros with compression:
[root]# mt -f /dev/st0 compression 1
[root]# dd if=/dev/zero of=/dev/nst0 bs=1000000 count=10000
10000000000 bytes (10 GB) copied, 100.147 seconds, 99.9 MB/s
gzip for comparison:
On a 1.6GHz Xeon box:
[root]# dd if=/dev/zero bs=1000000 count=10000 | gzip -c | wc -c
10000000000 bytes (10 GB) copied, 135.79 seconds, 73.6 MB/s
gives a compression ratio of 1030:1
And on a 2.9GHz Xeon box:
[root]# dd if=/dev/zero bs=1000000 count=10000 | gzip -c | wc -c
10000000000 bytes (10 GB) copied, 76.6102 seconds, 131 MB/s
Cheers,
Terry
-------------------------------------------------------------------------
This SF.Net email is sponsored by the Moblin Your Move Developer's challenge
Build the coolest Linux based applications with Moblin SDK & win great prizes
Grand prize is a trip for two to an Open Source event anywhere in the world
http://moblin-contest.org/redirect.php?banner_id=100&url=/
_______________________________________________
Bacula-users mailing list
Bacula-users AT lists.sourceforge DOT net
https://lists.sourceforge.net/lists/listinfo/bacula-users
|
