[Beowulf] Consumer vs. Enterprise Hard Drives in Clusters

[Bill Broadley]

Erwan Velu wrote:
> Bill Broadley wrote:
>> I found my actual results from a WD consumer 320GB drive (single platter) with
>> the drive on the bench vs inside a 1U node:
>>
>> http://cse.ucdavis.edu/bill/wd320-2MB-points.png
>>
>> http://cse.ucdavis.edu/bill/wd320-2MB-points-vibration.png
>>
>> Once the WD320 (a consumer drive) was replaced with a seagate es2 (raid
>> edition drive):
>>
>> http://cse.ucdavis.edu/bill/seagate-es.png
>>   
> 
> Hey Bill,
> Are we sure the only difference between both cases are only the vibrations ?

Nothing else changed.

> Do you use the same disk controller in all cases ?

Same physical node, not just the same kind of node.

> Does the workload is exactly the same ?

Yes, an idle node with a PXE installed image runs my benchmark.  My benchmark
is a very simple loop, basically read 4MB, print how long it took, repeat.
I'm using the raw device (/dev/sda), so the file system isn't installed.  This
was repeated over several nodes I had a cluster that was being built.  We
swapped out the disks and saw the dramatic difference.

I wish I had compared WD raid edition vs WD consumer, but I didn't have the
disks handy.

> Does the interconnect is the same ?

Yes, but I don't see how that could effect disk performance, nothing was going
over the network.

> 
> 
> Cheers,