Measuring ZFS configuration differences

Dan Langille dan at langille.org
Tue Nov 24 18:13:24 UTC 2015


> On Nov 18, 2015, at 4:12 PM, Dan Langille <dan at langille.org> wrote:
> 
> 
>> On Nov 17, 2015, at 9:31 PM, Marcelo Araujo <araujobsdport at gmail.com> wrote:
>> 
>> 2015-11-18 10:23 GMT+08:00 Dan Langille <dan at langille.org>:
>> 
>>> 
>>>> On Nov 17, 2015, at 8:46 PM, Marcelo Araujo <araujobsdport at gmail.com>
>>> wrote:
>>>> 
>>>> 2015-11-18 3:14 GMT+08:00 Dan Langille <dan at langille.org>:
>>>> 
>>>>> 
>>>>> On Nov 12, 2015, at 1:30 AM, Marcelo Araujo <araujobsdport at gmail.com>
>>>>> wrote:
>>>>> 
>>>>> 
>>>>> 
>>>>> 2015-11-12 6:34 GMT+08:00 Dan Langille <dan at langille.org>:
>>>>> 
>>>>>> On Oct 12, 2015, at 1:00 PM, Dan Langille <dan at langille.org> wrote:
>>>>>>> 
>>>>>>> Following up on the discussions during EuroBSDCon 2015 (Stockholm)
>>>>>> during the FreeBSD Developer
>>>>>>> Summit regarding various ZFS configuration settings, I write to start
>>>>>> our implementation phase now that some
>>>>>>> usual suspects have joined the list.
>>>>>>> 
>>>>>>> re https://wiki.freebsd.org/201510DevSummit/Performance
>>>>>>> 
>>>>>>> I think the first order of business is granting access rights to the
>>>>>> server (varm) in question:
>>>>>>> 
>>>>>>> http://dan.langille.org/2015/07/19/varm/
>>>>>>> 
>>>>>>> During the workshop, mention was made of serial access.  I can arrange
>>>>>> that.
>>>>>>> 
>>>>>>> The server has IPMI, however, my first thought:
>>>>>>> 
>>>>>>> 1 - connect a USB-serial cable to varm & link that to another server
>>> in
>>>>>> my rack.
>>>>>> 
>>>>>> Marcelo: At EuroBSDCon, was it you who mentioned a particular
>>>>>> configuration for the test machine which made
>>>>>> it easy to configure and run tests?  Was it PXE booting or something?
>>>>>> 
>>>>>>> 2 - create a jail in that server and give it access to that serial
>>>>>> connection
>>>>>>> 3 - redirect incoming port XYZ to that jail via a public-key-only ssh
>>>>>> connection
>>>>>>> 4 - give people access
>>>>>>> 
>>>>>>> Any suggestions?
>>>>>> 
>>>>>>>>>>>> Dan Langille
>>>>>> http://langille.org/
>>>>>> 
>>>>>> 
>>>>> Hello Dan,
>>>>> 
>>>>> Yes, was me :)
>>>>> 
>>>>> I mention about zopkio test framework.
>>>>> I gave a presentation last weekend at PyCon Hong Kong about it.
>>>>> 
>>>>> Here is my slides:
>>>>> 
>>> http://www.slideshare.net/araujobsd/functional-and-scale-performance-tests-using-zopkio
>>>>> 
>>>>> The good of Zopkio is, we can write tests at once and run it as much as
>>> we
>>>>> want in different machines. Also Zopkio depends of Naarad, that can
>>> parse a
>>>>> CSV file and create metrics and SLA over those metrics, plot graphs and
>>> so
>>>>> on. Pretty nice tool!!!
>>>>> 
>>>>> I'm wondering if we could start to test something and maybe show it at
>>>>> AsiaBSDCon and BSDCon(Canada) next year? What do you think?
>>>>> What I need right now would be a list of tests that we want to perform
>>> as
>>>>> well as what parameters we would like to take as metrics to compare.\
>>>>> 
>>>>> 
>>>>> For tests, we can start with this list:
>>>>> https://github.com/dlangille/zfs_benchmarks/issues
>>>>> 
>>>>> We can start as soon as I figure out how to provide access to the
>>>>> testers.  See above re serial connection.
>>>>> 
>>>>> I want to provide access, but I want to keep access restricted to only
>>>>> this box and not to the rest of my home LAN.  I plan to do this via a
>>>>> VLAN.
>>>>> 
>>>>> I could fire up a Rasperberry Pi and allow ssh into that.  Will that be
>>>>> enough
>>>>> power for what you need to do?
>>>>> 
>>>>> 
>>>> First of all, thanks to share the tests cases.
>>>> 
>>>> If I use zopkio, the best would be access SSH direct to the target
>>> machine
>>>> where I need to run the tests. For zopkio, I need to have my SSH KEY on
>>> the
>>>> target machine.
>>> 
>>> I am OK with this.
>>> 
>>>> As I don't know your network, maybe what you could do is: Via
>>> RasperBerry,
>>>> forward the SSH to the target machine, I will pass-through via your
>>>> RasperBerry where you can control the access for the rest of your LAN.
>>>> 
>>>> Another approach could be, two different subnets and a firewall. Or as
>>> you
>>>> said, VLANS.
>>> 
>>> I will be doing VLANS, which have yet to be set up.
>>> 
>>> The target system will have ZFS pools can be configured for different
>>> tests (i.e. raidz2 vs raidz3).
>>> This will involve gpart etc because the drives & pools will need to be
>>> 'wiped' between different test
>>> runs.
>>> 
>>> I seem to recall someone suggesting PXE boot and configuring the system
>>> remotely.  Does anyone
>>> recall that?  That aspect of the discussion was not recorded:
>>> https://wiki.freebsd.org/201510DevSummit/Performance
>>> 
>>> 
>> Bapt@ mentioned that, this is the way how we are doing in another project.
>> But in my point of view, it is not a must for our case!
>> 
>> The PXE wold be good if we try to test different of OS flavors, or build
>> different images.
> 
> OK.  The only thing holding us back is:
> 
> - adding the air filters to the case
> - moving to the new switch with the new VLANs
> 
> It's now a matter of time.

Done.

The server is ready for access.

Please send me your ssh-key off-list and I will create a login for you.

We do have some data on the system:

$ zpool list
NAME            SIZE  ALLOC   FREE  EXPANDSZ   FRAG    CAP  DEDUP  HEALTH  ALTROOT
benchmarking   21.8T  6.44T  15.3T         -    11%    29%  1.00x  ONLINE  -
music            30T  5.58T  24.4T         -     9%    18%  1.00x  ONLINE  -
random_mirror  2.72T   372K  2.72T         -     0%     0%  1.00x  ONLINE  -
zroot           220G  1.79G   218G         -     0%     0%  1.00x  ONLINE  -

The music pool has about 1.4TB of music files for testing via copy.

— 
Dan Langille
http://langille.org/







More information about the freebsd-performance mailing list