Hardware assisted parity computation - is it now worth it?

Hardware assisted parity computation - is it now worth it?

[Burn Alting]

Last year, there were discussions on this list about the possible
use of a 'co-processor' (Intel's IOP333) to compute raid 5/6's
parity data.

We are about to see low cost, multi core cpu chips with very
high speed memory bandwidth. In light of this, is there any
effective benefit to such devices as the IOP333?

Or in other words, is a cheaper (power, heat, etc) cpu with
higher memory access speeds, more cost effective than a
bridge/bus device (ie hardware) solution (which typically
has much lower memory access speeds)?

Burn Alting
burn@goldweb.com.au

Re: Hardware assisted parity computation - is it now worth it?

[Gordon Henderson]

On Thu, 13 Jul 2006, Burn Alting wrote:

> Last year, there were discussions on this list about the possible
> use of a 'co-processor' (Intel's IOP333) to compute raid 5/6's
> parity data.
>
> We are about to see low cost, multi core cpu chips with very
> high speed memory bandwidth. In light of this, is there any
> effective benefit to such devices as the IOP333?
>
> Or in other words, is a cheaper (power, heat, etc) cpu with
> higher memory access speeds, more cost effective than a
> bridge/bus device (ie hardware) solution (which typically
> has much lower memory access speeds)?

Something else to ponder: Is it worth it in terms of extra hardware
complexity, and additional software to drive it vs. the marginal increase
in speed it might give, when 99.9% of the time you are going to be
waiting on the external disk reading or writing the data compared to the
time it takes to generate the parity?

Personally, I'd opt for hardware simplicity over a marginal speed increase
anyday... However with todays energy prices on the increase, then it might
be something to consider, if it was going to give a significn power
savings, even so I'm not sure...

Gordon

Re: Hardware assisted parity computation - is it now worth it?

[Dan Williams]

On 7/13/06, Burn Alting <burn@goldweb.com.au> wrote:
> Last year, there were discussions on this list about the possible
> use of a 'co-processor' (Intel's IOP333) to compute raid 5/6's
> parity data.
The MD patches have been posted for review, and the hardware offload
pieces are nearing completion.

> We are about to see low cost, multi core cpu chips with very
> high speed memory bandwidth. In light of this, is there any
> effective benefit to such devices as the IOP333?
It is true that upcoming server platforms have an abundance of CPU
cycles, but what about the case where an IOP is the host processor?
This is the primary target of the current work.  Also, what about the
more expensive RAID6 conditions (2-failed disks) where there might be
benefits to having MD split its work over many CPUs?

Regards,

Dan

Re: Hardware assisted parity computation - is it now worth it?

[Burn Alting]

On Thu, 2006-07-13 at 14:16 -0700, Dan Williams wrote:
> On 7/13/06, Burn Alting <burn@goldweb.com.au> wrote:
> > Last year, there were discussions on this list about the possible
> > use of a 'co-processor' (Intel's IOP333) to compute raid 5/6's
> > parity data.
> The MD patches have been posted for review, and the hardware offload
> pieces are nearing completion.
> 
> > We are about to see low cost, multi core cpu chips with very
> > high speed memory bandwidth. In light of this, is there any
> > effective benefit to such devices as the IOP333?
> It is true that upcoming server platforms have an abundance of CPU
> cycles, but what about the case where an IOP is the host processor?

Meaning? - the kernel thread processes the write up to the generation of
parity then it calls/initiates another thread to compute parity/etc and
then context switches waiting for a soft interrupt from the other
'parity compute' thread?

> This is the primary target of the current work.  Also, what about the
> more expensive RAID6 conditions (2-failed disks) where there might be
> benefits to having MD split its work over many CPUs?

Agreed. Hiving off the parity generation to some (perhaps persistent)
thread that would compute parity for you in the most effective way (ie
multiple cpu's) is a good idea.

My main question is not so much the re-working of the code to make
better use of CPU's but is there is a future for IOP's which have a much
slower memory bandwidth than that of the cpu cores. As I understand it
we will soon be looking at a 5 to 1 speed difference!

> 
> Regards,
> 
> Dan

Burn

Re: Hardware assisted parity computation - is it now worth it?

[Bill Davidsen]

Burn Alting wrote:

>Last year, there were discussions on this list about the possible
>use of a 'co-processor' (Intel's IOP333) to compute raid 5/6's
>parity data.
>
>We are about to see low cost, multi core cpu chips with very
>high speed memory bandwidth. In light of this, is there any
>effective benefit to such devices as the IOP333?
>  
>

Was there ever? Unless you're running on a really slow CPU, like 386, 
with a TB of RAID attached, and heavy CPU load, could anyone ever see a 
measureable performance gain? I haven't seen any such benchmarks, 
although I haven't looked beyond reading several related mailing lists.

>Or in other words, is a cheaper (power, heat, etc) cpu with
>higher memory access speeds, more cost effective than a
>bridge/bus device (ie hardware) solution (which typically
>has much lower memory access speeds)?
>
An additional device is always more complex, and less tunable than a CPU 
based solution. Except in the case above where there is very little CPU 
available, I don't see much hope for a cost (money and complexity) 
effective non-CPU solution.

Obviously my opinion only.

-- 
bill davidsen <davidsen@tmr.com>
  CTO TMR Associates, Inc
  Doing interesting things with small computers since 1979