Gluster users - Sep 2014 - IOPS and CPU cycles, what is the relationship between the two?

Hello List,

The question is a fundamental for understanding storage and not 100% 
about GlusterFS but I fell like it's the right place to ask it.

I am trying to understand CPU cycles vs IOPS.
Per each IO There must be a number of cycles of the CPU.
So in order to choose the right CPU couple factors are needed to be 
right for the task.
For example a standard 7.2k RPS disk have maximum IOPS throughput.
A SSD drive will have twice and more IOPS throughput.
So let say I will take a celeron CPU with 2 cores which is a very low 
power consumer and I will put 10-20 Disks on this machine, Will it be 
able to handle the load of about 1.5k IOPS?
How can I calculate the maximum IOPS per CPU\machine possible?
There sure to be taken in account the Network traffic IOPS and also the 
whole network handling code using up cycles etc.

I do not have the answer and I am seeking for either a reference which 
will help to to understand it or a written explanation to the whole subject.

Thanks In Advance,
Eliezer Croitoru

On Mon, 08 Sep 2014 02:03:47 +0300
Eliezer Croitoru <eliezer at ngtech.co.il> wrote:
> The question is a fundamental for understanding storage and not 100% 
> about GlusterFS but I fell like it's the right place to ask it.
> 
> I am trying to understand CPU cycles vs IOPS.
> Per each IO There must be a number of cycles of the CPU.
It's not that much:
https://en.wikipedia.org/wiki/Direct_memory_access
> So in order to choose the right CPU couple factors are needed to be 
> right for the task.
> For example a standard 7.2k RPS disk have maximum IOPS throughput.
> A SSD drive will have twice and more IOPS throughput.
Usually at least more than factor 50, usual SATA disk with ~ 150 IOPS,
SSDs starting with ~ 7.500 IOPS (until 70k IOPS also in consumer disks
and above 100k IOPS in server equipment).
> How can I calculate the maximum IOPS per CPU\machine possible?
Take a look at the main board and the CPU where the connections between
IO (disk, network) and CPU are. They all have a specific bandwith.

Additionally you have max IOPS on your local discs, max IOPS on your
storage controller. You have to find the bottle neck yourselfs.
> There sure to be taken in account the Network traffic IOPS and also
> the whole network handling code using up cycles etc.
Handling the protocol layer may fill a CPU but TCP-Offloading exists to
make that better. What you want to do is to learn system
engineering. ;-)

Best Regards
Oli
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL:
<http://supercolony.gluster.org/pipermail/gluster-users/attachments/20140908/5f916721/attachment.sig>