RE: [xml-dev] When parsing speed matters (was Re: [xml-dev] No XML Binar

[noah_mendelsohn@us.ibm.com]

Michael Kay wrote:

> [Noah Mendelsohn wrote:]
> 
> > The analogy I use is to the CPUs in your printer: 
> 
> But the economics are rather different, surely? Printers are 
manufactured by
> the million. For more specialized functions, you have a much lower 
volume
> and therefore a higher unit cost and therefore you need a much more
> compelling story in terms of user benefits. Unless you're selling it 
like
> perfume: it's expensive so it must be good.

Yes, the economics are in general different, at least for many printers, 
though I doubt any but the most popular printers are manufactured in the 
millions.  I would expect the higher cost (e.g. wide roll) printers to be 
not too far off the price points and volumes of some XML accelerator 
boxes, but that's not really my point.  I'm not trying to make the case 
that because the numbers come out right for printers they necessarily do 
for XML.  I was making a qualitative statement:  among the reasons to 
consider having an outboard box is not necesesarily that it performs 
better per CPU cycle, but because it may be an economical way to add 
parallel CPU cycles when adding to your main processing units is starting 
to look expensive. 

Overall, it will be very interesting to see how these things play out as 
the exponential growth in the speed of individual processor cores starts 
to level off.  As I understand it, the projections are that transistor 
density will continue to grow well in coming years, but the ability to 
drive those transistors to higher GHz is being limited by the power draw 
that comes from higher frequencies, thinner gates, etc.  As Dave Patterson 
recently put it (in an absolutely terrific talk on future trends in 
computer architecture -- slides at [1]):

"Conventional Wisdom (CW) in Computer Architecture:
1. Old CW: Power is free, but transistors expensive
   New CW Power is expensive, but transistors are “free”
   Can put more transistors on a chip than have the power to turn on"

...and...

10. Old CW: Increasing clock frequency is primary
    method of performance improvement
    New CW: Processors Parallelism is primary
    method of performance improvement

11. Old CW: Don’t bother parallelizing app,
    just wait and run on much faster sequential
    computer
    New CW: No one building 1 processor per chip
            End of La-Z-Boy Programming Era"

In short, the CPU designers are running out of gas.  Single cores won't be 
getting much faster, but it will be cheaper to get lots more of cores, or 
to use the free transistors for other things.  You see this trend in the 
evolution from single to dual to quad cores in mainstream CPUs.  So, it 
will get increasingly tempting to do work like XML parsing and decryption 
in parallel with other activities.  When to do it on spare general purpose 
cores on your main chip, vs. on outboard general purpose boxes (which will 
have lots of cores too) vs. in specialized functions in either place will 
be interesting to watch in coming years.

Noah

[1] 
http://www.stanford.edu/class/ee380/Abstracts/070131-BerkeleyView1.7.pdf

--------------------------------------
Noah Mendelsohn 
IBM Corporation
One Rogers Street
Cambridge, MA 02142
1-617-693-4036
--------------------------------------