I am curious whether Kathryn McKinley's Power Wall argument is measuring a 
first order or second order effect. (This argument is essentially a furthering 
of Google's cost per killowatt hour utility company perspective, which says 
that as a process becomes more parallel it becomes more power hungry.)

It was only 2002 when a paper was published about the significant benefits of 
using control theory to scale up the DTM package on a chip; previous methods 
used peak power as an approximation for controlling thermal dissipation. These 
heuristics also throttled EVERYHING down, not just say instruction prefetching. 
I guess what I am trying to say is, from an outsider's perspective, the trade 
off might simply be based on current engineering practice rather than long term 
trend.

The other thing to note is that while compiler optimizations do play a role in 
power vs speed tradeoffs, higher level languages capable of describing 
transformations to themselves will play a bigger role.

Sent from my Droid Charge on Verizon 4GLTE

------Original Message------
From: David Corking <[email protected]>
To: "Fundamentals of New Computing" <[email protected]>
Date: Tuesday, December 20, 2011 9:28:05 AM GMT+0000
Subject: Re: [fonc] History of computing talks at SJSU

John Zabroski wrote:

> We have also yet to put into practice languages which limit the client
> run-time of an algorithm on a server (assuming the client can
> parameterize over the server's service in some disciplined way).
>
> Solving this problem will eliminate virtually all IT jobs.

Thanks for being provocative. In my turn, I think everything but your
last sentence is correct.

Google engineers have said that a key design parameter for their
services is the cost of instructions in kilowatt-hours. So the problem
may be less about idle cycles than about wasted cycles such as context
switches, inefficient algorithms and compiler optimization.

Meanwhile, I suspect servers for interactive services are rarely more
than 90% idle.

So I would suggest that efficient scheduling and request run-time
limits are important economic and environmental problems to solve, but
the solutions are unlikely to eliminate many digital tech jobs, as it
is unlikely to benefit society by more than the equivalent of two
Moore's law doubling cycles. That is a huge prize, but not the end of
the world as we know it.

My 2 farthings.
David

_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc


_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc

Reply via email to