Re: Dragonfly and Hyperthreading....

[EM1897@xxxxxxx]

In a message dated 2/21/2005 1:46:18 PM Eastern Standard Time, Matthew Dillon 
<dillon@xxxxxxxxxxxxxxxxxxxx> writes:

>
>:Ah, but with hyperthreading the performance boost comes at zero
>:cost. The difference in cost between a 3.6Ghz system and "something
>:faster" is perhaps 50%.People pay $300 more for a 3.6Ghz xeon than
>:a 3.4Ghz. A small gain has significant value. And dual cores will be very 
>:expensive for a long time. Intel and AMD aren't making them to 
>:cannibalize their own business; they're doing it to have something to
>:make their margins on.
>
>    My experience with regards to the computer industry is that people
>    pay a lot of money on stupid things, and this is one of those stupid
>    things.  There are arguably a few applications that can make use 
>    of it, but we are talking perhaps 0.01% (guess) of commercial
>    installations actually need the HT that they have.  The economics just
>    aren't there any more... vendors put such a high premium on HT and MP
>    technology that you can in fact get far better performance buying more
>    less powerful machines rather then one big honking machine, assuming
>    your application set can be partitioned that way.  Most of today's
>    applications cannot and that might seem to imply that HT is still good
>    for them, but the plain fact of the matter is that cpus are so powerful
>    these days they are not usually the bottleneck any more.  The real
>    issue in today's high performance computing world is cpu power per watt
>    consumed rather then raw cpu power, and since multi-core is capable
>    of providing far higher performance per watt of power then HT, that
>    makes HT a dead duck.
>
>    And I don't think dual-core is going to be expensive for very long,
>    not with AMD aiming it squarely at the consumer market and both AMD
>    and Intel likely moving in a direction that will result in having
>    *ONLY* multi-core chips in their lineup within the next 2-3 years.
>
>:Matt, would you comment on the following article:
>:
>:http://www-128.ibm.com/developerworks/linux/library/l-htl/
>:
>:benchmarks are usually misleading, but at least its not some kid in his
>:basement (probably not) doing this one.
>
>    Hyperthreading works best when you have a threaded application which is 
>    doing a mix of integer and memory operations, and works worst
>    when you have an application that is heavily integer, heavily memory,
>    or moderately to heavily FP oriented.  The biggest problem with HT
>    is that performance is extremely assymetric.  You could waste hundreds
>    of man hours writing an HT aware scheduler and still not have it work
>    well for all situations.
>
>    Even more to the point, HT technology was most useful in past years
>    when instruction pipelines were not able to make full use of the cpu's
>    resources.  That has changed as well.  Today's cpu's instruction 
>    pipelines are able to utility a far higher percentage of the cpu's
>    available resources, making HT a lot less useful.  HT has only been
>    viable up till now because Intel's cpu architecture sucks rocks compared
>    with AMD's, but Intel can no longer compete by boosting clock rate so
>    now they are stuck... they have to slow their cpus down and make them
>    more efficient, and that will kill HT's effectiveness even if you 
>    ignore the heat issue.

I think .01% is a bit low. But then again, 99.9% of all applications are 
UP, and no-one seems interested in supporting that market???

I'm most entertained by my customers who buy 3.4Ghz EE CPUS and use
32-bit/33Mhz NICs in their servers....

Most of the stuff going on now is just plain PR, because as you say, most 
people don't need more than
what today's CPUS deliver anyway. I've never seen any compelling evidence
that AMD-based machines are faster in practice.  Right now I'm testing a 
3.4/800Ghz P4 system with 1K cache and PCIExpress and its slower than 
a 3.06/533 Xeon with 512K cache and PCI-X. The faster CPU is faster in raw 
processing, but the "architecturally slower" machine is faster with larger 
packets, which should be the opposite.  The truth is that no-one has any 
idea what they're buying, and you can't even use the "specs" to infer that 
one machine is faster than another because there are so many external 
factors.

My problem is that "2-3 years" is a lifetime, and I have no confidence that 
things will play out the way we suspect. What to do while Dragonfly and
FreeBSD are in transition? Retire? Switch to LINUX? Argh!

On a side note, do you see intel moving their Pentium-M technology to
the "desktop"? My 2.0Ghz P-M notebook kicks butt.