[mythtv-users] Do you build mythtv from source, and why?

[Raymond Wagner]

On 3/30/2011 08:27, Brian Long wrote:
> On Tue, Mar 29, 2011 at 3:07 PM, Raymond Wagner <raymond at wagnerrp.com 
> <mailto:raymond at wagnerrp.com>> wrote:
>
>
>     The Atom processor is the perfect example of the wrong product at the
>     right time.  Netbooks had just started coming out, and while some ran
>     Linux, most people wanted to stick with Windows, which requires an x86
>     processor.  The ULV mobile processors are considerably higher
>     performance than the Atoms while in the same power envelope, but
>     due to
>     all the crap Intel ripped out of the design, the Atoms were low
>     transistor count and dirt cheap to manufacture.  ARMs at the time had
>     similar performance, with a small fraction of the power
>     consumption, but
>     since they couldn't run Windows, the Atom won out.  Windows 8 is
>     expected out sometime next year with ARM support.  I don't expect the
>     Atom line to survive much beyond that.
>
>
> It appears some companies are betting the farm on Atom surviving a 
> little longer:
> http://www.seamicro.com/?q=node/38

Seamicro is only riding the existing PR with that one.  As mentioned, 
the ARMs have about as much integer performance as the Atom, at a 
fraction of the power consumption and cost.  While the ARMs have poor 
floating point performance, the Atoms aren't especially good at that either.

So let's look at this thing.  512 1.66GHz cores in a 10U chassis, 
replacing 40U of traditional server.  Now a traditional server processor 
is going to do around 3-4 times the work per core as these Atoms, and 
since you can stuff 32 cores into a 2U server, you're looking at the 
same 8-12U of equipment as this thing, not the 40U they claim.

Now what about price?  That 2U chassis with 8 quad core 2.93GHz Xeons, 
192GB of memory, and 12 500GB drives is going to run just under $10K.  
For somewhere between $40K-$60K, you can build an equivalent traditional 
server system to that Seamicro box.  The Seamicro box costs $150K.

Now what about power consumption?  That 2U chassis is going to eat up 
around 1kW under full load, or maybe half that while idle.  While each 
of those 256 Atoms is only going to consume around 8W each, figure 
double that when the memory and chipset is taken into account.  That 
puts it at maybe 4kW for the whole system under full load.  Now here's 
the problem, those desktop Atom chips don't really idle.  One of the 
things Intel ripped out in the name of cheapness was the clock stepping 
and power gating.  That Atom will consume less power under load, but 
more power while idle, than the traditional server.  Surely for one 
third the price, that 20% less power consumption under load is not 
something you would be all that concerned about.

What other concerns are there with going with that Atom solution?  
First, you have software licensing.  Commercial server software often 
charges per processor or per core.  Often times the software costs 
considerably more than the server, meaning it's advantageous to get the 
power powerful per thread system you can.  The Atoms lose out big time 
here.  Next, you have inter-connectivity.  The Atom will have 64 gigabit 
ethernet ports, compared to the traditional server's 40.  With typical 
web serving applications, that's plenty.  For most HPC applications, 
that's inadequate, and you're going to want something fast with low 
latency.  To get infiniband, you're looking at an extra $15K for 20 
ports on the traditional server, plus another $10K for a switch to 
handle them.  For the Atoms, it's simply not an option and you're out of 
luck.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.mythtv.org/pipermail/mythtv-users/attachments/20110330/86f1cfb0/attachment.html