There have been some studies recently about what kind of system would best run Oracle and serve transactions. (See the Alpha-multiprocessor article in this year's ISCA proceedings.)
The answer is a multiprocessor (because Oracle is threaded) with a shared L2 cache (because the threads share most of the L2 footprint.) The advantage is quite large.
So, what IBM is building here is a server engine. Which is why there will be a 128-processor system, but there won't be Altivec.
So saith the article. Well, it would be nice if
we could trust their vigilance. But look at
http://www.joeyskaggs.com/
Joey has suckered nationwide TV news shows, over
and over, for decades. Joey says it's not getting
harder. When he was showing CNN his "legal
supercomputer" that had "judged O.J. Simpson was
guilty", one of the cameramen said something
like "Hey, haven't I been here before? Something
about fish condos?" But the anchorman didn't
notice and didn't check. And that's typical.
describes a system that I saw demoed in, um, 1993. The system computed an image at 10-15 frames per second, and each pixel of the image had a brightness proportional to the distance. I stood in front of the camera, and a human-shaped outline appeared on the screen. I put my hand out, and that part of the image got brighter, with my hand brigher than my arm.
The algorithm took several hundred megaflops. Lots of today's chips should be able to run that at a useful frame rate.
Back in the 1970's, Jean Sammett published a book with the Tower of Babel on the cover. It listed several thousand (yes, thousand) compilers. Her criteria was that the compiler had to have been used by at least one person other than its author.
By now, the total has to be past 10,000. Of course, almost all of them were abandoned, and the authors have moved on. But dozens-to-hundreds have active user communities. Poke around a bit.
So, GCC is simply the most famous of the free compilers. And, since it's GPLed, anyone who wants to can start up their own variant of GCC.
"Someone who owns a newspaper can pick up the phone to editor and say 'run headlines I like'."
And someone who is the biggest advertiser in a magazine can cancel all his advertising, to punish them for printing negative things about his products.
Microsoft has actually done this. Is there freedom of the (trade) press when someone has a monopoly? No.
ENIAC: Important but not at all seminal
on
ENIAC Story on NPR
·
· Score: 1
ENIAC is historically important. But it wasn't the first digital electronic system - that was either Colussus or the ABC, depending on how you count. And the chief designer of the ENIAC had studied the ABC.
And, ENIAC wasn't the first stored-program computer. It was the generation after ENIAC that did that. Maurice Wilkes (who is still with us) was the first to run a stored program. Good random access memory came even later: Wilkes (and the ENIAC follow-on, the Univac I) had to use serial memory.
And ENIAC had a very inefficient design. It didn't use binary arithmetic: it sent the digit "7" by sending 7 pulses, like an old mechanical telephone exchange. It was probably Von Neumann who realised that binary would use a lot fewer vaccuum tubes.
So, the ENIAC guys were bold, and pretty good, but they don't have a clear claim to just about any of the Big Concepts. What they really contributed was to snag the interest and committment of a broad military and scientific community.
From least important to most important: (1) code (2) data (3) how you will tell if it works (4) interfaces (5) motherhood
Motherhood is things like: is it supposed to be fast? Or small? Or intuitive to radio astronomers? Or prototyped by Friday? When push comes to shove, what goals win? This is the crucial stuff.
Interfaces are the next most important, because chunks that use well chosen interfaces can be reimplemented without causing a disturbance in the force.
Unfortunately, the "subtle properties" aren't consist from chip to chip. That is, the chip manufacturer guarantees that certain properties are the same from chip, and this search process wound up using properties that weren't in the set.
Slashdot Mirror
There have been some studies recently about what kind of system would best run Oracle and serve transactions. (See the Alpha-multiprocessor article in this year's ISCA proceedings.)
The answer is a multiprocessor (because Oracle is threaded) with a shared L2 cache (because the threads share most of the L2 footprint.) The advantage is quite large.
So, what IBM is building here is a server engine. Which is why there will be a 128-processor system, but there won't be Altivec.
So saith the article. Well, it would be nice if
we could trust their vigilance. But look at
http://www.joeyskaggs.com/
Joey has suckered nationwide TV news shows, over
and over, for decades. Joey says it's not getting
harder. When he was showing CNN his "legal
supercomputer" that had "judged O.J. Simpson was
guilty", one of the cameramen said something
like "Hey, haven't I been here before? Something
about fish condos?" But the anchorman didn't
notice and didn't check. And that's typical.
http://www.cs.cmu.edu/afs/cs.cmu.edu/user/webb/htm l/iccv-stereo.html
describes a system that I saw demoed in, um, 1993. The system computed an image at 10-15 frames per second, and each pixel of the image had a brightness proportional to the distance. I stood in front of the camera, and a human-shaped outline appeared on the screen. I put my hand out, and that part of the image got brighter, with my hand brigher than my arm.
The algorithm took several hundred megaflops. Lots of today's chips should be able to run that at a useful frame rate.
http://www.cs.cmu.edu/People/wearable/index.html
and click on the "Digital Ink" button.
No kidding. Lots of Linux software is already hard to see on 1600x1280 screens. Not to mention web pages - why don't browsers have a Zoom feature?
Most compiler projects give the results away.
Back in the 1970's, Jean Sammett published a book with the Tower of Babel on the cover. It listed several thousand (yes, thousand) compilers. Her criteria was that the compiler had to have been used by at least one person other than its author.
By now, the total has to be past 10,000. Of course, almost all of them were abandoned, and the authors have moved on. But dozens-to-hundreds have active user communities. Poke around a bit.
So, GCC is simply the most famous of the free compilers. And, since it's GPLed, anyone who wants to can start up their own variant of GCC.
"Someone who owns a newspaper can pick up the phone to editor and say 'run headlines I like'."
And someone who is the biggest advertiser in a magazine can cancel all his advertising, to punish them for printing negative things about his products.
Microsoft has actually done this. Is there freedom of the (trade) press when someone has a monopoly? No.
ENIAC is historically important. But it wasn't the first digital electronic system - that was either Colussus or the ABC, depending on how you count. And the chief designer of the ENIAC had studied the ABC.
And, ENIAC wasn't the first stored-program computer. It was the generation after ENIAC that did that. Maurice Wilkes (who is still with us) was the first to run a stored program. Good random access memory came even later: Wilkes (and the ENIAC follow-on, the Univac I) had to use serial memory.
And ENIAC had a very inefficient design. It didn't use binary arithmetic: it sent the digit "7" by sending 7 pulses, like an old mechanical telephone exchange. It was probably Von Neumann who realised that binary would use a lot fewer vaccuum tubes.
So, the ENIAC guys were bold, and pretty good, but they don't have a clear claim to just about any of the Big Concepts. What they really contributed was to snag the interest and committment of a broad military and scientific community.
From least important to most important:
(1) code
(2) data
(3) how you will tell if it works
(4) interfaces
(5) motherhood
Motherhood is things like: is it supposed to be fast? Or small? Or intuitive to radio astronomers? Or prototyped by Friday? When push comes to shove, what goals win? This is the crucial stuff.
Interfaces are the next most important, because chunks that use well chosen interfaces can be reimplemented without causing a disturbance in the force.
Unfortunately, the "subtle properties" aren't consist from chip to chip. That is, the chip manufacturer guarantees that certain properties are the same from chip, and this search process wound up using properties that weren't in the set.