BTW, before I encountered kdb, my favorite db was PG. The comparison between it and Oracle is very legitimate. It seems PG is specifically targeting Oracle users and looks increasingly like Oracle to me. Let's see. In PG, you can write your stored procedures in TCL. PG/SQL or Perl. In Oracle, only in PL/SQL (The thing with Java is vaporware) The object extentions really do work in PG. In Oracle, they are again vaporware. (PG invented the 'object-relational' genre and has major headstart there. Two decades or some such)
Everything that is possible in Oracle is possible in a whole lot of other databases, only with a fraction of administrative overhead. I am no fan of Microsoft's but even SQL Server is better (and way cheaper too) There is too much crud in Oracle. The whole thing should've been rewritten 20 years ago. People buy Oracle because they are lemmings. If you must buy your Oracle, try Oracle rdb. They don't market it much, but it is way better than their main product.
Now, if you really want a database that will blow anything else out of the water, try kdb (www.kx.com, and no I don't work for them, but I promise you will be amazed)
Crystal is the most powerful stimulant I know of and is also a major aphrodisiac (this is why tweakers watch porn all the time:) Apparently, there is also a drug called 'Ice' or 4-Metaminarex (sp?) which is just as strong as crystal, but only lasts 6-8 hours. I did not try it, but I find it would rule to have something as good as crystal that doesn't keep you up for 3 days for a change.
I find coke to be not addictive at all too. All the (claimed) coke-fiends I know switched to speed when they tried it. A lot more bang for the buck. Personally, I cannot understand how can one be addicted to something with such a mild effect. Speed is surely very addictive, but if you can afford it, who cares?
I agree, coke is fucking weak. For me, it only lasts half an hour. Coffeine is stronger, but I get all jittery on higher dose/day 2. (I took it in pills, cause I hate the taste of coffee) Crystal is best - you are up for up to 3 days (it goes downhill after that), your head is crytal-clear (pun intended) and you get tons of work done due to all that extra noradrenaline. And there are indeed plently of people who can keep their crack habit under control, contrary to popular belief. Ephedrine is OK, very much like mild speed, but gram for gram it is a stronger cardiovascular stimulant than speed (ie ups your heart rate and temperature to larger extent) Useful to burn some fat, but may not be what you want. You want mental stimilation, not physical, right?
This is old hat. The <A HREF="http://magnet.consortia.org.il/ConSolar/Sab<nobr>i<wbr></wbr></nobr> n/Zas/Zas3.html">Israeli solar chimney</A> is built high rather than wide, and on top of reducing CO2, also generates electricity! Check it out.
From what I hear from bike nuts, the lighter and stiffer the frame, the better. Hence, Al frames are better than gimmicky Ti ones. (Ti is a relatively soft metal) It would be interesting to see more applications that take advantage of its extreme resistance to compression (outside of Russian subs that is)
I always found it amazing how easy it is to buy a congressman in US and how difficult to buy a DMV clerk. I would say US is very non-corrupt at the lowest levels of government and very corrupt at the top. It is customary for continental Western European states to be other way around. I guess this is because the (usually) socialist governments don't have a cozy relationship with big business. I currenly live in Austria. Here it is customary to bring a gift when you go for appointment for official business (cash is out though) Cops sometimes takes bribes. Have you ever considered that a bit of corruption can be useful though, as a tool against government opression? For example, I find most speed limits to be way on the low side and I like fast driving, so I find it advantageous to be able to pay my way out of any traffic violation. US is quite unique in that it has relatively few opressive regulations, but the ones that are there are actually actively enforced, no matter how stupid.
I was recently searching for 'tetrachromicity' on Google and Google went like, 'Do you mean tetrachomity'? And indeed, that was the correct term I was looking for. That was quite impressive.
I think there is a place for another engine a-la Alta Vista that doesn't try to be smart about things but does keyword matches very well. It's really good for those cases when you remember nothing from the site except an exact sentence or a few words exactly as they come in the text. I like the norwegian fastsearch thingie for that even better than Alta Vista - it may be an even larger version of the same thing.
Ask Jeeves is OK, too. Of course, the natural language processing doesn't work, but it's pretty good as metasearch. Apparently, they get the best stats and the most useful inputs because when people try to enter fully qualified English queries, they simply end up giving them more search terms then usual.
'bourgeois.xenote.com' (The company was french. It also effectively prevented my boss (an ex-engineer) to log in and tinker with my config too often - the French can't spell it either:) I liked 'morningwood.interval.com' also.
Best flaimebait:
'makethenetfreeforchildpornography.c2.net'
Bext convention for small machine groups:
1. Unpack the case and check out the 'made in' sticker.
2. Name the machine 'stupid white man' in the languguage of that country, eq 'gringo', 'baca-gajin', etc.
I don't see how the 29000 is any special, in particular relating to Tao. It seems to be a rather generic chip, quite a me-too effort from AMD. Not that it was bad, but let's see:
Unbounded register set' is most commonly achieved by register renaming these days. SPARC-style sliding 'register windows' are generally agreed to have been a mistake due to implementation complexity (note how SPARC has always been behind on Mhz to competing RISC designs (as long as those still existed) Lots of registers are good, but don't forget that 128 of them takes up 8 bits of space per-register from the opcode. The first (almost) orthogonal register set I am aware of is the one of M68K, which was quite CISC. Harvard architecture takes away the option of writing self-modifying code, which can be quite powerful. Every chip since 68020 is Harvard, so the point is moot anyway. I totally agree on CCs - the architectures than left them out (MIPS was first) lost a lot more than the miniscule gains in implementation complexity. Not only should you have CCs, but also predication (ala ARM and now Merced), IMHO. I think the best-ever RISC architecture was ARM, followed by PowerPC, but in general, I find the concept of a 'register' a bit offensive. Come to think of it, a 'register' is just memory with addressing restrictions and is just a performance hack. My ideal architecture would be some kind of a vector IRAM (say Playstation 2?), I guess.
This URL: http://www-2.cs.cmu.edu/~koopman/stack_computers/c ontents.html Contains a cool book on stack machines. Check it out before you slam them. Having said that, there is no good reason for the JVM to be stack- rather than register based. I think the best way to have a portable runtime were to leave the code at the stage of an AST and let every VM implementer to compile to what they want (including native) Tao 'portable assembly' thingie is way kewl and has been around for a while, too. (from since way before Java, etc.)
This URL: http://www-2.cs.cmu.edu/~koopman/stack_computers/c ontents.htmlContains a cool book on stack machines. Check it out before you slam them. Having said that, there is no good reason for the JVM to be stack- rather than register based. I think the best way to have a portable runtime were to leave the code at the stage of an AST and let every VM implementer to compile to what they want (including native)
Tao 'portable assembly' thingie is way kewl and has been around for a while, too. (from since way before Java, etc.)
I remember how back in 1995 another intern at SGI spent the whole summer trying to figure out how to program the blinkenlights on big SGI/Crays only to find out that they are, in fact, generated by a separate battery-powered plastic unit generating essentually random patterns. It looked like something out of a toystore. Apparently, the serious government customer expects blinkenlights. It aint a supercomputer without the blinkenlights. Is it CM/War games legacy? The same guy previously did successefully write a library to control blinkenlights on a 20-CPU Sequent, though.
K is mostly written in itself and you can freely examine that code from inside the environment. There is a C part to it, but I am not sure you'll be able to make any sense out of it if you are not used to APL-like code density. To put it simply, the C part of K is written in K implemented in C macro language. This was generative programming way before it was officially invented. To get a taste of what I am talking about, look at the source of A+ at
(A+ is a precursor language to K, that Arthur wrote at Morgan Stanley. It is a lot more APL-like than K)
K has been known to beat carefully written C code on a few ocassions and here are some possible explanation (It is weird, considering that K is itself written in C)
a) K uses bulk allocation. The smallest memory unit it uses is a VM page. To achieve the same result with C, you will have to hoist all your mallocs to the upmost level and do all your own memory allocation - this is not a very common thing you see.
b) A lot of C code uses stdio here and there. K's native I/O mechanism is memmap (This is another reason why the Bell labs benchmark rubs K the wrong way - it uses stdio (and loops like Jason said before))
c) Do not underestimate the power of lambda/delayed eval combined with JIT compiler. For example, it is possible for the K runtime to detect an inverse function to a given function and drop both before evaluation.
d) K tends to produce code that is nearly branchless, eq not a single explicit loop in the whole of kdb! This is very nice on your memory subsystem. In general, kdb's got to be world's only IN-CACHE db (at 150K, the whole thing tends to stay cache resident at all time. I checked - the code had virtually no trace. The data trace is extremely linear due to the embracing of inherent temporal ordering in relational data. This is very unusual - for example Oracle is all over memory and routinely trashes 4MB caches. How is that for a paradigm shift if I tell you that relational databases are actually among the most vectorizable/paralleralizable codes out there?!? They ran the wrong software on those Crays...
K was invented way before OSS really got off. kdb is free for non-commercial use and their license is no more anal than say MySQL's was before recently while the quality is way above. I hope K will get used, even if only for writing kdb extensions/stored procs. You have no idea how useful a FUNCTIONAL database can be (as opposed to relational)
It is released in binary, free for non-commercial use:
Arthur told me that he will open-source it, if there is sufficient interest. It is not very difficult to obtain the source from kx for your private use right now.
It should be possible to translate Scheme to K code 1-to-1, but there is usually a better way to do it in K that is not entirely obvious. If you entire code is so short that you can have the whole thing in front of your eyes, you can really get a lot of optimizing the code on global scale - it is kind of similar to how the hardware guys look for a logic function that just happens to produce the same result as the required gate. kdb tends to SHRINK with subsequent releases, as it adds features. You can think of K conceptually as a sort-of-Scheme that is based on a triple
(hashvalue, value, next) instead of a usual double (first, next) Unlike any LISP I am aware about, there are a lot of features oriented towards system programming specifically and the performance is exemplary - it takes very non-trivial COMPILED C code to beat INTERPRETED K, especially on large datasets.
kx never had any official TPC scores for kdb submitted, (lots of work for such a small company) but the issue of performance vs. Oracle or even research IN-MEMORY databases such as Blitz is not even funny. I did millions of transactions per second to a STANDARD IDE DRIVE (using a custom transaction reordering scheme and multidimensional bulk transactions. You can't do either in Oracle) Fundamentally, the performance of kdb is gated solely by the throughput of hardrive or network controller (whichever one is smaller)
At some point, the 'line noise' is worth it. Would you rather have a line of noise than 10 'readable' lines? What about 100? 1000? There is a break-even point somewhere. I had 4! lines of K and KSQL replacing 2.5K lines of Java. (132-wide lines, not 80, mind you) It is just amazing how little fluff K has. Take this example:
convert seconds to hours, min, sec format:
24 60 60 _vs numseconds
convert IP address to octets:
4#8 _vs ipaddress
Same function ('vector to scalar', basically a conversion to variable radix) is used for both and it is a function that is part of the language and is used all over the place. Pretty much elsewhere, both tasks would have you use completely different calls into some sort of standard library. Result - code bloat.
K is the most elite language out there, bar none, IMHO. Check it out.
I have seen Intel confirming elsewhere that their Hyperthreading is based on SMT design they got with Alpha, which is in turn based on Washington's SMT. I went to an SMT talk by Washington guys while at Berkeley - very elegant and impressive stuff indeed.
BTW, this project is two years old - you hear about it only now because the IAEA has hired a PR consultancy to put a spin on the story. (This comes from an IAEA insider source)
Slashdot Mirror
BTW, before I encountered kdb, my favorite db was
PG. The comparison between it and Oracle is very legitimate. It seems PG is specifically targeting Oracle users and looks increasingly like Oracle to me. Let's see. In PG, you can write your stored procedures in TCL. PG/SQL or Perl. In Oracle, only in PL/SQL (The thing with Java is vaporware) The object extentions really do work in PG. In Oracle, they are again vaporware. (PG invented the 'object-relational' genre and has major headstart there. Two decades or some such)
Everything that is possible in Oracle is possible in a whole lot of other databases, only with a fraction of administrative overhead. I am no fan of Microsoft's but even SQL Server is better (and way cheaper too) There is too much crud in Oracle.
The whole thing should've been rewritten 20 years ago. People buy Oracle because they are lemmings.
If you must buy your Oracle, try Oracle rdb. They don't market it much, but it is way better than their main product.
Now, if you really want a database that will blow anything else out of the water, try kdb (www.kx.com, and no I don't work for them, but I promise you will be amazed)
Up until Vietnam, they used a lot of meth for just that purpose.
Crystal is the most powerful stimulant I know of and is also a major aphrodisiac (this is why tweakers watch porn all the time :) Apparently, there is also a drug called 'Ice' or 4-Metaminarex (sp?) which is just as strong as crystal, but only lasts 6-8 hours. I did not try it, but I find it would rule to have something as good as crystal that doesn't keep you up for 3 days for a change.
I find coke to be not addictive at all too. All the (claimed) coke-fiends I know switched to speed when they tried it. A lot more bang for the buck. Personally, I cannot understand how can one be addicted to something with such a mild effect.
Speed is surely very addictive, but if you can afford it, who cares?
I agree, coke is fucking weak. For me, it only lasts half an hour. Coffeine is stronger, but I get all jittery on higher dose/day 2. (I took it in pills, cause I hate the taste of coffee) Crystal is best - you are up for up to 3 days (it
goes downhill after that), your head is crytal-clear (pun intended) and you get tons of work done due to all that extra noradrenaline. And there are indeed plently of people who can keep their crack habit under control, contrary to popular belief. Ephedrine is OK, very much like
mild speed, but gram for gram it is a stronger cardiovascular stimulant than speed (ie ups your heart rate and temperature to larger extent) Useful to burn some fat, but may not be what you want. You want mental stimilation, not physical, right?
Grr, of course, there should be no space in the URL: http://magnet.consortia.org.il/ConSolar/Sabi n/Zas/Zas3.html
This is old hat. The <A HREF="http://magnet.consortia.org.il/ConSolar/Sab<nobr>i<wbr></wbr></nobr> n/Zas/Zas3.html">Israeli solar chimney</A> is built high rather than wide, and on top of reducing CO2, also generates electricity! Check it out.
From what I hear from bike nuts, the lighter and
stiffer the frame, the better. Hence, Al frames are better than gimmicky Ti ones. (Ti is a relatively soft metal) It would be interesting to
see more applications that take advantage of its
extreme resistance to compression (outside of Russian subs that is)
Try REI. I am also a proud owner of their titanium fork and spoon kit 'for alpinists'. :)
I always found it amazing how easy it is to buy a congressman in US and how difficult to buy a DMV clerk. I would say US is very non-corrupt at the lowest levels of government and very corrupt at the top. It is customary for continental Western European states to be other way around.
I guess this is because the (usually) socialist governments don't have a cozy relationship with big business. I currenly live in Austria. Here it is customary to bring a gift when you go for appointment for official business (cash is out though) Cops sometimes takes bribes. Have you ever
considered that a bit of corruption can be useful though, as a tool against government opression?
For example, I find most speed limits to be way on the low side and I like fast driving, so I find it advantageous to be able to pay my way out of any traffic violation. US is quite unique in
that it has relatively few opressive regulations, but the ones that are there are actually actively enforced, no matter how stupid.
I was recently searching for 'tetrachromicity'
on Google and Google went like, 'Do you mean
tetrachomity'? And indeed, that was the correct
term I was looking for. That was quite impressive.
I think there is a place for another engine a-la
Alta Vista that doesn't try to be smart about things but does keyword matches very well. It's
really good for those cases when you remember nothing from the site except an exact sentence
or a few words exactly as they come in the text.
I like the norwegian fastsearch thingie for that
even better than Alta Vista - it may be an even
larger version of the same thing.
Ask Jeeves is OK, too. Of course, the natural
language processing doesn't work, but it's pretty
good as metasearch. Apparently, they get the best
stats and the most useful inputs because when
people try to enter fully qualified English queries, they simply end up giving them more search terms then usual.
Best singles:
:)
'bourgeois.xenote.com' (The company was french. It
also effectively prevented my boss (an ex-engineer) to log in and tinker with my config too often - the French can't spell it either
I liked 'morningwood.interval.com' also.
Best flaimebait:
'makethenetfreeforchildpornography.c2.net'
Bext convention for small machine groups:
1. Unpack the case and check out the 'made in'
sticker.
2. Name the machine 'stupid white man' in the languguage of that country, eq 'gringo', 'baca-gajin', etc.
I don't see how the 29000 is any special, in particular relating to Tao. It seems to be a rather generic chip, quite a me-too effort from
AMD. Not that it was bad, but let's see:
Unbounded register set' is most commonly achieved by register renaming these days. SPARC-style sliding 'register windows' are generally agreed to have been a mistake due to implementation complexity (note how SPARC has always been behind on Mhz to competing RISC designs (as long as those still existed) Lots of registers are good, but don't forget that 128 of them takes up 8 bits
of space per-register from the opcode. The first (almost) orthogonal register set I am aware of is the one of M68K, which was quite CISC. Harvard architecture takes away the option of writing self-modifying code, which can be quite powerful. Every chip since 68020 is Harvard, so the point is moot anyway. I totally agree on CCs - the architectures than left them out (MIPS was first) lost a lot more than the miniscule gains in implementation complexity. Not only should you have CCs, but also predication (ala ARM and now Merced), IMHO. I think the best-ever RISC architecture was ARM, followed by PowerPC, but in
general, I find the concept of a 'register' a bit offensive. Come to think of it, a 'register' is
just memory with addressing restrictions and is just a performance hack. My ideal architecture would be some kind of a vector IRAM (say Playstation 2?), I guess.
Lots of registers
This URL: http://www-2.cs.cmu.edu/~koopman/stack_computers/c ontents.html Contains a cool book on stack machines. Check it out before you slam them. Having said that, there is no good reason for the JVM to be stack- rather than register based. I think the best way to have a portable runtime were to leave the code at the stage of an AST and let every VM implementer to compile to what they want (including native) Tao 'portable assembly' thingie is way kewl and has been around for a while, too. (from since way before Java, etc.)
This URL: http://www-2.cs.cmu.edu/~koopman/stack_computers/c ontents.htmlContains a cool book on stack machines. Check it out before you slam them. Having said that, there is no good reason for the JVM to be stack- rather than register based. I think the best way to have a portable runtime were to leave the code at the stage of an AST and let every VM implementer to compile to what they want (including native)
Tao 'portable assembly' thingie is way kewl and has been around for a while, too. (from since way before Java, etc.)
I remember how back in 1995 another intern at SGI spent the whole summer trying to figure out how to program the blinkenlights on big SGI/Crays only to find out that they are, in fact, generated by a separate battery-powered plastic unit generating essentually random patterns. It looked like something out of a toystore. Apparently, the serious government customer expects blinkenlights. It aint a supercomputer without the blinkenlights. Is it CM/War games legacy? The same guy previously did successefully write a library to control blinkenlights on a 20-CPU Sequent, though.
K is mostly written in itself and you can freely examine that code from inside the environment. There is a C part to it, but I am not sure you'll be able to make any sense out of it if you are not used to APL-like code density. To put it simply, the C part of K is written in K implemented in C macro language. This was generative programming way before it was officially invented. To get a taste of what I am talking about, look at the source of A+ at
http://www.aplusdev.org/
(A+ is a precursor language to K, that Arthur wrote at Morgan Stanley. It is a lot more APL-like than K)
K has been known to beat carefully written C code on a few ocassions and here are some possible explanation (It is weird, considering that K is itself written in C)
a) K uses bulk allocation. The smallest memory unit it uses is a VM page. To achieve the same result with C, you will have to hoist all your mallocs to the upmost level and do all your own memory allocation - this is not a very common thing you see.
b) A lot of C code uses stdio here and there. K's native I/O mechanism is memmap (This is another reason why the Bell labs benchmark rubs K the wrong way - it uses stdio (and loops like Jason said before))
c) Do not underestimate the power of lambda/delayed eval combined with JIT compiler. For example, it is possible for the K runtime to detect an inverse function to a given function and drop both before evaluation.
d) K tends to produce code that is nearly branchless, eq not a single explicit loop in the whole of kdb! This is very nice on your memory subsystem. In general, kdb's got to be world's only IN-CACHE db (at 150K, the whole thing tends to stay cache resident at all time. I checked -
the code had virtually no trace. The data trace is extremely linear due to the embracing of inherent temporal ordering in relational data. This is very unusual - for example Oracle is all over memory and routinely trashes 4MB caches. How is that for a paradigm shift if I tell you that relational databases are actually among the most vectorizable/paralleralizable codes out there?!? They ran the wrong software on those Crays...
I screwed up the link in the previous post, try this one:
here
K was invented way before OSS really got off. kdb is free for non-commercial use and their license is no more anal than say MySQL's was before recently while the quality is way above. I hope K will get used, even if only for writing kdb extensions/stored procs. You have no idea how useful a FUNCTIONAL database can be (as opposed to relational)
It is released in binary, free for non-commercial use:
Arthur told me that he will open-source it, if there is sufficient interest. It is not very difficult to obtain the source from kx for your private use right now.
It should be possible to translate Scheme to K code 1-to-1, but there is usually a better way
to do it in K that is not entirely obvious. If you entire code is so short that you can have the whole thing in front of your eyes, you can really
get a lot of optimizing the code on global scale - it is kind of similar to how the hardware guys look for a logic function that just happens to produce the same result as the required gate. kdb tends to SHRINK with subsequent releases, as it adds features. You can think of K conceptually as a sort-of-Scheme that is based on a triple
(hashvalue, value, next) instead of a usual
double (first, next) Unlike any LISP I am aware about, there are a lot of features oriented towards system programming specifically and the performance is exemplary - it takes very non-trivial COMPILED C code to beat INTERPRETED K, especially on large datasets.
kx never had any official TPC scores for kdb submitted, (lots of work for such a small company) but the issue of performance vs. Oracle or even research IN-MEMORY databases such as Blitz is not even funny. I did millions of transactions per second to a STANDARD IDE DRIVE (using a custom transaction reordering scheme and multidimensional bulk transactions. You can't do either in Oracle) Fundamentally, the performance of kdb is gated solely by the throughput of hardrive or network controller (whichever one is smaller)
At some point, the 'line noise' is worth it.
Would you rather have a line of noise than 10
'readable' lines? What about 100? 1000? There is a break-even point somewhere. I had 4! lines of K and KSQL replacing 2.5K lines of Java. (132-wide lines, not 80, mind you) It is just amazing how little fluff K has. Take this example:
convert seconds to hours, min, sec format:
24 60 60 _vs numseconds
convert IP address to octets:
4#8 _vs ipaddress
Same function ('vector to scalar', basically a conversion to variable radix) is used for both and it is a function that is part of the language and is used all over the place. Pretty much elsewhere, both tasks would have you use completely different calls into some sort of standard library. Result - code bloat.
K is the most elite language out there, bar none, IMHO. Check it out.
I have seen Intel confirming elsewhere that their
Hyperthreading is based on SMT design they got
with Alpha, which is in turn based on Washington's
SMT. I went to an SMT talk by Washington guys while at Berkeley - very elegant and impressive
stuff indeed.
This BBC article has a video, which also ran
e ws id_1824000/1824576.stm
on BBC news for a couple of days:
http://news.bbc.co.uk/hi/english/world/africa/n
BTW, this project is two years old - you hear
about it only now because the IAEA has hired
a PR consultancy to put a spin on the story.
(This comes from an IAEA insider source)
Most Rado watches are made from ultra-hard
ceramics also: www.rado.com
Metals made transparent by photonic layer structure:
http://eetimes.com/story/OEG19991108S0095
This is much more useful than transparent armor,
IMHO, if it can indeed be applied to photonic
band-gap filtering...