Re: Got idle CPU cycles?

From: Uncle Al (UncleAl0_at_hate.spam.net)
Date: 09/09/04 

Date: Thu, 09 Sep 2004 14:02:25 -0700

"P.T. Breuer" wrote:
>
> Uncle Al <UncleAl0@hate.spam.net> wrote:
> > 1) It isn't written to network. The program eats 98+% of all
> > available CPU cycles. It is incredibly CPU-intensive. It would
> > freeze a network solid.
>
> Err .. cpu cycles do not influence networks. Tasks that do i/o get
> priority, so your thread will leave the kernel whenever anyone wants to
> talk. On a premptive kernel you won't even get the 1/1000th second
> latency born of finishing the time slice, it'll leave immediately.
>
> Plus it's probably your thread doing the networking, so I don't see how
> you figure that.

That sounds OK, but it doesn't necessarily work real world. Linux
handles task swapping crisply, Windows congeals into sluggish mush.
Any host CPU will have its pipes completely filled. There won't be
any excess capacity to do other things. User latencies will appear.
This thing filled a cluster of 16 Opteron-848s and ran it flat out.
One suspects folks would be happier were it explicitly confined to a
single CPU. Anway, it's crunching in four Xeons, two Athlons, and a
Pentium4 as you read this. Intel makes crap CPUs.

> > 2) It is also a small project. We've got 17,000 points
> > calculated.
>
> Ponits of what? Some graph? Don't you have an analytic expression? It's
> a sphere, no?

Build a sphere, calculate CHI. Increment the radius, do it again.
The smallest radial increment at which CHI fluctuates varies inversely
as the square of the radius. That the original sampling of 711 radii
between 3 A and 1.1 million A radius worked at all floored a number of
people. Even with 25K-50K points we are looking at an infinitesimal
fraction of the available data. Do you believe that random extremes
neatly cancel in a random sampling?

We have three analytic expressions: Pure math that sets the slope
(looks to be exact vs. real world), a one parameter model for the
intercept (within 6% of real world for quartz), and grinding out real
world data by the book. It is the goodness of fit between theory and
crystal structure data that we are exploring. The crystal structure
is only good to 3% (which is awesome. It's a tough row to hoe). A
real world crystal will contain a remarkable assortment of
imperfections. What is truth? What is publishable?
 
> > on silly. The far right points - the rightmost one being 444
> > quadrillion atoms - chewed 900 hrs in a cluster of 16 Opteron-848s at
>
> Oh, you want to change the size of the sphere.

The quartz lattice is anisotropic. Every time the radius increases
enough to add two atoms (has to be at least two by symmetry) CHI
fluctuates. We need an extrapolated answer for a centimeter diameter
lump from calculations at atomic lattice scales. There is a certain
justified belief that crap will happen while we are bridging the
distance gap. It turns out things look clean over six orders of
magnitude no matter how close we look. That silences learned critics
(who are familiar with their own fast and loose antics).
 
[snip]

> You can label the lines red and blue.

I'll use tiny crosses instead of tiny dots against total disappearance
with photoreduction, and reverse the line black to white over the
dense area. "8^>) Gnuplot can only mark the axes in whole pixels.
Referees will complain about an extra pixel or one missing here and
there. You can imagine the dialog,

Referee, "Get an graphic artist to do it."
Uncle Al, "Plot 25,000 points to four significant figures by hand?"

Where would we be without Management? 

-- 
Uncle Al
http://www.mazepath.com/uncleal/
 (Toxic URL! Unsafe for children and most mammals)
http://www.mazepath.com/uncleal/qz.pdf

Relevant Pages

  • Re: 2003 Server slowed to a crawl
    ... click Processes then CPU column to sort by ... Investigate a possible Network problem (is there a broadcast storm ... >> method for telling the computer's processor that it needs attention. ... >> which have signed drivers (such hardware is sold with a Microsoft Windows ...
    (microsoft.public.windows.server.general)
  • Re: Game Company- Java Server Thread Priority
    ... > Program takes 0% of the CPU load, and when processing a turn goes up to ... An example of network waits. ... expected it to go *up* (since the server box wasn't running the client too). ... involved a network access which slowed it down massively. ...
    (comp.lang.java.programmer)
  • Re: network bitrate of a poll of processes
    ... I would like to know the bitrate of a pool of child processes that use ... a network connection, how can I have something like netstat -w1 ... CPU load b) IO on drivec) network load ... PID %MEM %CPU PPID TT STAT TIME COMMAND ...
    (freebsd-hackers)
  • Re: 2003 Server slowed to a crawl
    ... click Processes then CPU column to sort by processes ... > the network, are backing up large amounts of user data during production ... > method for telling the computer's processor that it needs attention. ... > which have signed drivers (such hardware is sold with a Microsoft Windows ...
    (microsoft.public.windows.server.general)
  • Re: a proposed callout API
    ... of the network implementation in such different operating systems ... The basic premise is that you can make better use of any available cpu ... the operating systems are so different from each other I don't think ...
    (freebsd-arch)