30 May
2011
30 May
'11
4:15 a.m.
10 GOTO 10
Parallelize THAT.
By the 1970's many timesharing and batch type compilers would kick out the
embarrassingly useless programs without even trying (the environment
well tested by millions of college student's programming attempts).
I think the line between embarrassingly useless and embarrassingly parallel
may not always be obvious. In many university-type batch processing systems,
student programs that ran for more than 5 CPU seconds were summarily
terminated as "obviously being stuck in some kind of infinite loop that the compiler
can't catch yet". Just a few
years later the same problems probably qualified for Grand Challenge grants :-)
I remember when a very big fear of the US govt, was that the Japanese would
wipe us out with Fifth Generation computers. (The WMD of the 80's)
Boy, did that miss the mark. The Japanese came up with the Fuji Eagle (perhaps
the most defining commodity-computing storage element ever) and today
teraflop-class vector processors are in every fanboy's computer and
they're called "graphics cards".
Tim.
30 May
30 May
8:27 a.m.
On 30 May 2011 at 7:15, Shoppa, Tim wrote:
Boy, did that miss the mark. The Japanese came up with
the Fuji Eagle
(perhaps the most defining commodity-computing storage element ever)
and today teraflop-class vector processors are in every fanboy's
computer and they're called "graphics cards".
The big enemy of supercomputer engineering today is Moore's Law. By
the time design and testing is complete and product is rolling off
the line, it's obsolete. Better to spend your time building a fast
bit of silicon and hoping that it'll see a 6 month to one year
product window.
Going through the effort of designing an building a mainframe-style
supercomputer such as a Cray 3 would be folly today. Witness Steve
Chen's "box full of Pentiums" as how badly this sort of thing can go.
And it's getting steadily worse. Designing and implementing a new
CPU architecture today is becoming a matter of how fast one can cook
up a chunk of VHDL.
Eventually, I suspect we'll get to the point where a chip will be
obsolete by the time the trade rags have gotten aroud to announcing
it.
Consider that people were prosecuted for espionage for passing
"secrets" to the Soviets about the Saxpy Matrix-1 CPU in 1987. Who
even remembers Saxpy now (aside from the name of a LINPACK program
set)? Saxpy made their product announcement in 1987 and they were
pretty much out of business by the end of 1988.
I can't even keep up with CPU and microcontroller announcements
today.
--Chuck
9:34 a.m.
On 5/30/11 11:27 AM, Chuck Guzis wrote:
The big enemy of supercomputer engineering today is
Moore's Law. By
the time design and testing is complete and product is rolling off
the line, it's obsolete. Better to spend your time building a fast
bit of silicon and hoping that it'll see a 6 month to one year
product window.
As far as semiconductor process evolution goes, definitely. The type
of innovation that can make an impact and last a while is architectural.
Going through the effort of designing an building a
mainframe-style
supercomputer such as a Cray 3 would be folly today. Witness Steve
Chen's "box full of Pentiums" as how badly this sort of thing can go.
Pentia? ;)
-Dave
--
Dave McGuire
Port Charlotte, FL
2:05 p.m.
On 30 May 2011 at 13:35, Fred Cisin wrote:
On Mon, 30 May 2011, Chuck Guzis wrote:
Back to the "nine women, baby in one month" quip--now that I think of
it, I don't recall who was the first to say it. I know that I heard
it from a number of people prior to 1975, but it's also in Fred
Brooks' "Mythical Man month" book.
Does anyone know the origin of the statement? Brooks' book was
mandatory reading for project managers at CDC in the 70s. Is it
still used at all?
--Chuck
P.S. I believe that Gene Amdahl had something to say about
parallelizing problems...
Eventually, I suspect we'll get to the point
where a chip will be
obsolete by the time the trade rags have gotten aroud to announcing
it.
"To see the current chips, press F5" 
2:32 p.m.
On Mon, 30 May 2011, Chuck Guzis wrote:
Back to the "nine women, baby in one month"
quip--now that I think of
it, I don't recall who was the first to say it. I know that I heard
it from a number of people prior to 1975, but it's also in Fred
Brooks' "Mythical Man month" book.
Does anyone know the origin of the statement? Brooks' book was
mandatory reading for project managers at CDC in the 70s. Is it
still used at all?
I would be interested if you can find any authoritative attribution. I
have even heard it attributed to Werner Von Braun, but I don't think that
that is reliable.
It'll probably be as hard to find definitive attribution as it is for "It
is easier to get forgiveness than permission."
or:
"Those who wouldst take away our rights will start by redefining them as
privileges [of the crown]?"
--
Grumpy Ol' Fred cisin at xenosoft.com
10:02 p.m.
On 5/30/2011 4:32 PM, Fred Cisin wrote:
On Mon, 30 May 2011, Chuck Guzis wrote:
I've thought for years that
one was from Grace Hopper, but ICBW.
Later,
Charlie C.
Back to the "nine women, baby in one
month" quip--now that I think of
it, I don't recall who was the first to say it. I know that I heard
it from a number of people prior to 1975, but it's also in Fred
Brooks' "Mythical Man month" book.
Does anyone know the origin of the statement? Brooks' book was
mandatory reading for project managers at CDC in the 70s. Is it
still used at all?
I would be interested if you can find any authoritative attribution. I
have even heard it attributed to Werner Von Braun, but I don't think that
that is reliable.
It'll probably be as hard to find definitive attribution as it is for "It
is easier to get forgiveness than permission." or:
"Those who wouldst take away our rights will start by redefining them as
privileges [of the crown]?"
--
Grumpy Ol' Fred cisin at xenosoft.com
31 May
31 May
4:25 a.m.
On Monday (05/30/2011 at 08:27AM -0700), Chuck Guzis wrote:
Going through the effort of designing an building a mainframe-style
supercomputer such as a Cray 3 would be folly today. Witness Steve
Chen's "box full of Pentiums" as how badly this sort of thing can go.
So, there are some people who still want such things-- and some people
who will still build them.
I know a thing or two about this system,
http://www.nas.nasa.gov/hecc/resources/pleiades.html
as I did the architecture for the first version rolled out in 2007--
at 487 TFlop in 100 racks. They've since scaled out to 168 racks,
adding higher-end CPUs and can now cross the 1.0 PFlop line with LINPACK.
I'm no longer with SGI but this product line continues and has been
updated to QDR Infiniband, 10 GbE and other enhancements to the
interconnect.
I can assure you that the performance of the interconnect is still the
bottleneck ;-) which is one reason why we put two parallel IB fabrics
into this thing.
A node in this machine is pretty much a standard Intel Xeon based server
without any disks or power supply. It's built on a slightly different
PCB form-factor but is otherwise a standard reference design for such
a platform. The whole point was to leverage that commodity stuff so
that you could A LOT of it together and scale it out at lower cost than
totally proprietary stuff.
Chris
--
Chris Elmquist
9:46 a.m.
On 31 May 2011 at 6:25, Chris Elmquist wrote:
And I'm sure the spooks have stuff that would put your example to
shame. But where does the NASA setup stand in terms of the Peta-
(e.g. NCSA Blue Waters) and Tera-scale (too numerous to mention)
projects going on now? A few TFlop sure won't seem like much in one
or two years, if what I'm reading is any indication.
Used to be that you could get 5-10 years out of a supercomputer. Now
that notion seems quaint and will be even more strange in the future.
--Chuck
5:36 p.m.
On Tuesday (05/31/2011 at 09:46AM -0700), Chuck Guzis wrote:
On 31 May 2011 at 6:25, Chris Elmquist wrote:
And I'm sure the spooks have stuff that would put your example to
shame. But where does the NASA setup stand in terms of the Peta-
(e.g. NCSA Blue Waters) and Tera-scale (too numerous to mention)
projects going on now? A few TFlop sure won't seem like much in one
or two years, if what I'm reading is any indication.
well, it cruises at 750+ TFlop... and peaks at 1.2 PFlop so it's in the
ball park.
Spooky things are often not general purpose systems that can run anybody's
old MPI apps on a familiar OS. Although they can really scream, they
are purpose built for a single kind of problem and highly optimized for
that situation and as a result, either can't even run or really suck at
running less Spooky problems.
I don't think PFlop machines are going to be falling from trees any time
soon and many that do get there, also aren't general purpose. The thing
we got from using commodity hardware is that it also runs "commodity
software"... eg, SuSE and RedHat Linux right out of the box.
Things are happening with GPUs that improve the local to the node
performance but you've still got big challenges getting high node to node
bandwidth in order to really make it an MPP. Proprietary interconnects
are interesting and cool and maybe really fast but they are also really
expensive and tend to be one of the gating factors in the scalability.
We used off-the-shelf Infiniband which was (relatively) cheap and so
we could use more of it and scale out farther (and faster) for the
same money.
Used to be that you could get 5-10 years out of a
supercomputer. Now
that notion seems quaint and will be even more strange in the future.
Pleiades ran its first benchmarks in November 2007 and made #3 on Top500
then. As far as I know, it is still one of the primary resources for
NASA Ames and they continue to expand it and keep it fully utilized.
So, in 6 months, it'll be 5 years old.
Chris
--
Chris Elmquist
4720
days inactive
4722
days old
9 comments
6 participants
participants (6)
-
cclist＠sydex.com -
chrise＠pobox.com -
cisin＠xenosoft.com -
csquared3＠tx.rr.com -
mcguire＠neurotica.com -
tshoppa＠wmata.com