On Fri, 29 Apr 2011 13:39:29 +1000, <doiggl(a)velocitynet.com.au> wrote:
On Tue, 26 Apr 2011 17:06:04 -0400, Greg Freemyer
<greg.freemyer(a)gmail.com>
wrote:
On Tue, Apr 26, 2011 at 11:46 AM, David Greaves
<david(a)dgreaves.com>
wrote:
>> On 19/04/11 18:49, Greg Freemyer wrote:
>>>
>>> On Tue, Apr 19, 2011 at 1:28 PM, Stephen Shaw<sshaw(a)decriptor.com>
>>> wrote:
>>>>
>>>> On Sat, Apr 16, 2011 at 02:19,<doiggl@velocitynet.com.au> wrote:
>>>>>
>>>>> Hello
>>>>> With the opensuse build service [obs] getting more usage every day
>>>>> I wonder if this would make a diffence :-
>>>>>
>>>>> 1. Use solid state disks [ssd] to improve build times of
package
>>>>> rpms
>>>>> on, perhaps point the BuildRoot or _tmppath to the ssd's
>>>>>
>>>>> Reason for asking.
>>>>> Improve the packaging times during busy times on obs.
>>>>>
>>>>> Cheers Glenn.
>>>>
>>>> Unfortunately there is a cost associated, but it would be cool to
get
>>>> a bunch of fusion IO cards and see how
they perform.
>>>>
>>>> Here is a youtube video of 1 fusion IO card handling 4500 movies
>>>> streaming at once. If I understood it right.
>>>>
>>>>
http://www.youtube.com/watch?v=9J5xGwdmsuo
>>>>
>>>
>>> Clearly buying SSD costs money.
>>>
>>> But buying a $1000 250 GB SSD for one of the build servers would be
>>> cheaper than buying a whole new server.
>>>
>>> So the next time an expansion to the build farm is budgeted, I hope
>>> getting SSDs for the buildroot is considered as an option.
>>>
>>> If the thread from 6 months ago is still accurate, having a SSD might
>>> double the compile speed of a build farm node.
>>
>> Don't forget that workers are very, very write-heavy. This is not the
>> best
>> I/O pattern for SSD; and of course the vast majority of data written
>> during
>> a build-preparation is never re-read.
>
> Why is a write heavy load not good for a SSD? Performance or Lifetime?
>
> For performance, the big issue is random i/o vs sequential.
>
> A rotating disk and a SSD should have similar sequential speeds.
>
> But for random i/o the SSD should be much, much faster in that it
> should barely slow down due to a random i/o write pattern.
>
> Also, the mkfs version in 11.4 tells the SSD that all of the sectors
> in the partition are unused (trim'ed) prior to starting the format.
> That should ensure consistent highspeed operation out of the SSD. If
> the build farm is up to 11.4 yet, just the mkfs.ext4 package could be
> used I suspect.
>
> For lifetime, I don't know. But a typical build seems to take several
> minutes, so the writes to the SSD don't seem like they would be that
> outrageously heavy.
>
> Greg
More info from the company which did above demo
http://www.iomax.com.au/products.htm
See the links to the areas covered which covered
- 4,500 DVDs playing from a single ioDrive Octal.
- Running 512 Virtual machines running on one VMWare View host.
- Over 1,000 DVDs playing from a single ioDrive.
Also
http://iomax.com.au/performance.htm
Interesting indeed.
I wonder if one could be tried out ?
-Glenn
--
To unsubscribe, e-mail: opensuse-buildservice+unsubscribe(a)opensuse.org
For additional commands, e-mail: opensuse-buildservice+help(a)opensuse.org