Lots of software that typically Sun used to push on their platforms (over PCs that were much slower/less reliable in the past) is not as aggresively multithreaded as it should be. It didn't matter back then, and it matters now, but that would necessitate rewriting a lot of algorithmic code.
I'm sure the difference in speed would be much more even or in Suns' favor if the IC software could properly utilize the 8-way machine. (Have you considered running multiple jobs in parallel on the Sun an compiling throughput stats?)
It's stuff like Oracle that really shine on Sun boxes. But the stuff that tends to run on Solaris really good also tends to still run on PCs really good, because that's easy to do.
We were going to spend $15K about 3 years ago to upgrade an ailing E450 to max out proc and memory. We were supporting multiuser MATLAB/Simulink.
Instead, we threw that money at 6 dual Athlon XPs.
In 3 months, the E450 was only being used to run distributed.net. If a single box was given 2 jobs, it could complete them 225% faster than the Sun, and in the worse case, 150% faster in a contrived memory constrained situation.
Multiply by 6 and we easily more than tripled the capacity, while reducing overhead costs/maintenance.
Sigh. Sun was pissed at us too. We did this a number of times. PC hardware (if you make good choices) has caught up. What are you going to do?
The only problem I ever seem to have with Dells are their power supplies (and how adding any extra disks past configured capacity always seems to end up in a power supply dying).
I wish they would overspec them. Sure, they may not be as quiet, but jeez!
3d output is good for comprehension. (DUH!) All of which means that 3d interfaces are suitable when you need to see things that are 3-dimensional (double DUH!).
Note that I said big screens, but I didn't necessarily say 3D. A vertical dimension might be implied by stacking or scaling (ala OSX) if needed, but it doesn't need to be full blown 3D (this could become detremental).
And you wouldn't want to have "surfaces" presented at any oblique angle, especially when they can be occluded or if transparency is involved. (We are talking about search results, right?) It prevents you from reading any text or identifying icons at different parts of it with equal resolving power.
And forget about identifying physical objects in 3d. Having a space filling characteristic just limits your ability to present things because there's too many additional oppurtunities for "clutter" (you should see my desk!)
Anything 3d will immediately slow down your interaction to a snails pace as you manipulate your environment. Even if it was a virtual mind-meld into a matrix like environment... "walking" to your search result and activating would take longer than a quick scroll down a result list with text blurbs.
Intuitive does not mean good.
It should be efficient, and become good through acclimation. Just like riding a bicycle. It seems garish at first, but it makes perfect sense later on.
Just look at the interface from Minority Report. We should all be so lucky to have UIs like that. The answer is big screens, "front page snippets" representation of documents/results for at a glance viewing, and multidimensional arrangement where dimension (and tagging) is based on attributes (relevance, date, accuracy). Dimension could mean position in space or in a hierachy, etc.
everytime calls their product "holistic" I just want to take it and have at it with an awl and hammer. Then I say to them "Now it's DEFINITELY hole-tastic!"
Atoms are character classes (letters, ranges, or bracket expressions), conjunctions of said classes, or a paranthetized expression (like in maths).
You have two conjunctions. The first is concatentation is what you get when you put one atom right after another (they both have to appear in that order). The other is alternation (pipe) where either the left atom or right atom must appear.
Finally modifiers are an optional number of repetitions for each atom to match. The default is from 1 to 1 (exactly one). * means from 0 to infiinity, ? means 0 to 1, + means from 1 to infinity, and {x,y} means from x to y.
as every person who has spent the time to put a project on freshmeat or sourceforge probably believes he or she deserves commercial backing.
Meanwhile you can probably list on a piece of notebook paper all the open source software that has had the benefit of paid, full time employees.
Let's see: Apache, Linux, Eclipse, Gnome (but really that's Helix now), Mono, OpenOffice (at one time), Mozilla, KDE (QT), Octave, CSound, Mosix, linux-wlan, Slash, Bugzilla, Zope, (the AOL webserver written in TCL which I can't remember whatsit called...), anaconda, RPM, (and the redhat conf tools), blender, Cinerella, and that's about all that I can think of. About 40% of these are instances of software that was once closed that was subsequently opened.
Some projects get corporate support, but they don't get salaried employees. I'm not saying other software sucks, but a lot of the big-name stuff was not all a hobbiest effort, and this has to be recognized.
XP plays nice on an NT system where you may have many users who want to store their documents in more than one place, have their own clipart... hell just being able to RUN the software as an unprivledged user is kinda nice.
You can only put something in the public domain if you explicitly state so, or you remit your right to retain copyright.
That is, some thing with an (at the time) unattributed author is not considered public domain until such time the author declares his/her intent. Up until that point, you may treat it as public domain with the knowledge that at any time the author could come after you if they so choose (up to the time limitations, anyway).
So it's not really public domain unless you can prove it is. All it takes is one irate person on a project team who disagrees...
GPL helps to protect you from some of that. They can't "take it back" once it's released GPL, as long as you haven't monkeyed with it and rereleased it without attribution or source.
from a 32-bit address aligned on a 8-byte boundary without penalty if it's in the cache already (to an FP/MMX reg, of course). Access to two sequential 32-bit, 4 byte aligned values into GP registers can NOT be serialized, IIRC.
This is one of the reasons why you WANT to use MMX instructions to do simple processing of matricies or arrays or whatever... more efficient memory access.
They chose to decouple addressing logic and instruction logic. You need to use prefix bytes to use 64-bit/extended register ops. Thus you are free to mix 32-bit and 64-bit instructions as needed in a single program to save instruction cache. As a result, 64-bit instructions follow the same "formula" as 32-bit instructions. In a way this must have helped immensely in design and Q/A because all the prefix byte really does is turn off features that remap 16->8 and mask the upper 32-bits of GP regs. Mostly it's unchanged. The big thing was the integrated MMU (with 64-bit addressing support), and all the extra instructions to manage it.
solaris uses mmap exclusively (internally) to emulate ALL file system access (swap, local, nfs, you name it). This allows it to intelligent manage the page cache and disk cache with the ultimate flexibility.
Linux has a similar behavior for certain filesystems (it's fs dependant)
You have always been able to fetch 64-bits from RAM into a register as quickly as you could load a 32-bit register. You would use an MMX/FP instruction. The path from the cache to the P6 cpu is at least 64-bits wide, and cache lines (and memory bus) are as wide or wider.
Now they've extended that all the way to pointers and integers (and thus more ADDR lines for 64 (well 48-bit addressing), but that's a different story)
You can't do MMX and floating point in parallel. But you CAN do 64-bit arithmetic (which used to require MMX) and floating point in parallel, or 64-bit integer arithmetic and MMX/SSE/3dnow in parallel. This opens up the door for lots of additional speed optimizations by internal instruction level parallelism. Also it lets you do things with less instructions (and hence clocks) which used to be cumbersome before (faster fixed point/arbitrary precision math, large multiply/acc. operations, saturation math with 32-bit quantities)
but even when programming in assembler, it isn't so bad. In fact using the x86_64 IA assembler is pretty nice in long mode (AMD is happy to send you the manuals gratis, to help out). Not as symmetric as say MIPS, but it's getting better.
The IA is the IA. It has little to do with how anything else works (which is all modern at this point). The only pain is having non-fixed length opcodes, but we've solved that problem (TLBs) so we get to use instruction cache better. I don't really see the problem!
Slashdot Mirror
Nice eva. Can we just see the eva and screw the whiny pilots?
Lots of software that typically Sun used to push on their platforms (over PCs that were much slower/less reliable in the past) is not as aggresively multithreaded as it should be. It didn't matter back then, and it matters now, but that would necessitate rewriting a lot of algorithmic code.
I'm sure the difference in speed would be much more even or in Suns' favor if the IC software could properly utilize the 8-way machine.
(Have you considered running multiple jobs in parallel on the Sun an compiling throughput stats?)
It's stuff like Oracle that really shine on Sun boxes. But the stuff that tends to run on Solaris really good also tends to still run on PCs really good, because that's easy to do.
that uber-gaming boxes have largely surpassed the requirements of many workstations (with the exception of RAM requirements).
I mean, did you think the ATI Fire XL was anything but an R350 with 4 times the ram, thereby justifying it costing twice as much?
I didn't think so.
We were going to spend $15K about 3 years ago to upgrade an ailing E450 to max out proc and memory. We were supporting multiuser MATLAB/Simulink .
Instead, we threw that money at 6 dual Athlon XPs.
In 3 months, the E450 was only being used to run distributed.net. If a single box was given 2 jobs, it could complete them 225% faster than the Sun, and in the worse case, 150% faster in a contrived memory constrained situation.
Multiply by 6 and we easily more than tripled the capacity, while reducing overhead costs/maintenance.
Sigh. Sun was pissed at us too. We did this a number of times. PC hardware (if you make good choices) has caught up. What are you going to do?
The only problem I ever seem to have with Dells are their power supplies (and how adding any extra disks past configured capacity always seems to end up in a power supply dying).
I wish they would overspec them. Sure, they may not be as quiet, but jeez!
Just thought you should know, so you don't look like an asshole.
::sipping freshly brewed coffee::
3d output is good for comprehension. (DUH!)
All of which means that 3d interfaces are suitable when you need to see things that are 3-dimensional (double DUH!).
Note that I said big screens, but I didn't necessarily say 3D. A vertical dimension might be implied by stacking or scaling (ala OSX) if needed, but it doesn't need to be full blown 3D (this could become detremental).
And you wouldn't want to have "surfaces" presented at any oblique angle, especially when they can be occluded or if transparency is involved. (We are talking about search results, right?) It prevents you from reading any text or identifying icons at different parts of it with equal resolving power.
And forget about identifying physical objects in 3d. Having a space filling characteristic just limits your ability to present things because there's too many additional oppurtunities for "clutter" (you should see my desk!)
Anything 3d will immediately slow down your interaction to a snails pace as you manipulate your environment. Even if it was a virtual mind-meld into a matrix like environment... "walking" to your search result and activating would take longer than a quick scroll down a result list with text blurbs.
Intuitive does not mean good.
It should be efficient, and become good through acclimation. Just like riding a bicycle. It seems garish at first, but it makes perfect sense later on.
Just look at the interface from Minority Report. We should all be so lucky to have UIs like that. The answer is big screens, "front page snippets" representation of documents/results for at a glance viewing, and multidimensional arrangement where dimension (and tagging) is based on attributes (relevance, date, accuracy). Dimension could mean position in space or in a hierachy, etc.
everytime calls their product "holistic" I just want to take it and have at it with an awl and hammer. Then I say to them "Now it's DEFINITELY hole-tastic!"
There is a DEFINITE central structure.
Atoms, modifiers, and conjunctions.
Atoms are character classes (letters, ranges, or bracket expressions), conjunctions of said classes, or a paranthetized expression (like in maths).
You have two conjunctions. The first is concatentation is what you get when you put one atom right after another (they both have to appear in that order). The other is alternation (pipe) where either the left atom or right atom must appear.
Finally modifiers are an optional number of repetitions for each atom to match. The default is from 1 to 1 (exactly one). * means from 0 to infiinity, ? means 0 to 1, + means from 1 to infinity, and {x,y} means from x to y.
That's it.
Double letters are harder to "trip" over.
as every person who has spent the time to put a project on freshmeat or sourceforge probably believes he or she deserves commercial backing.
Meanwhile you can probably list on a piece of notebook paper all the open source software that has had the benefit of paid, full time employees.
Let's see: Apache, Linux, Eclipse, Gnome (but really that's Helix now), Mono, OpenOffice (at one time), Mozilla, KDE (QT), Octave, CSound, Mosix, linux-wlan, Slash, Bugzilla, Zope, (the AOL webserver written in TCL which I can't remember whatsit called...), anaconda, RPM, (and the redhat conf tools), blender, Cinerella, and that's about all that I can think of. About 40% of these are instances of software that was once closed that was subsequently opened.
Some projects get corporate support, but they don't get salaried employees. I'm not saying other software sucks, but a lot of the big-name stuff was not all a hobbiest effort, and this has to be recognized.
What are we arguing about, anyway?
XP plays nice on an NT system where you may have many users who want to store their documents in more than one place, have their own clipart... hell just being able to RUN the software as an unprivledged user is kinda nice.
n/t
Just look at the posting history... wow! That's called a reality check.
because a lot of the hard up-front work and design was funded by Netscape (AOL): they had full-time salaried programmers working on it.
The same goes with IBM and Eclipse.
Few projects will have this kind of backing and resources.
You can only put something in the public domain if you explicitly state so, or you remit your right to retain copyright.
That is, some thing with an (at the time) unattributed author is not considered public domain until such time the author declares his/her intent. Up until that point, you may treat it as public domain with the knowledge that at any time the author could come after you if they so choose (up to the time limitations, anyway).
So it's not really public domain unless you can prove it is. All it takes is one irate person on a project team who disagrees...
GPL helps to protect you from some of that. They can't "take it back" once it's released GPL, as long as you haven't monkeyed with it and rereleased it without attribution or source.
from a 32-bit address aligned on a 8-byte boundary without penalty if it's in the cache already (to an FP/MMX reg, of course).
Access to two sequential 32-bit, 4 byte aligned values into GP registers can NOT be serialized, IIRC.
This is one of the reasons why you WANT to use MMX instructions to do simple processing of matricies or arrays or whatever... more efficient memory access.
They chose to decouple addressing logic and instruction logic. You need to use prefix bytes to use 64-bit/extended register ops. Thus you are free to mix 32-bit and 64-bit instructions as needed in a single program to save instruction cache. As a result, 64-bit instructions follow the same "formula" as 32-bit instructions. In a way this must have helped immensely in design and Q/A because all the prefix byte really does is turn off features that remap 16->8 and mask the upper 32-bits of GP regs. Mostly it's unchanged. The big thing was the integrated MMU (with 64-bit addressing support), and all the extra instructions to manage it.
than a Barton clock for clock. Just as the Barton was more efficient per Hz than the Throughbred. :-)
solaris uses mmap exclusively (internally) to emulate ALL file system access (swap, local, nfs, you name it). This allows it to intelligent manage the page cache and disk cache with the ultimate flexibility.
Linux has a similar behavior for certain filesystems (it's fs dependant)
You have always been able to fetch 64-bits from RAM into a register as quickly as you could load a 32-bit register. You would use an MMX/FP instruction. The path from the cache to the P6 cpu is at least 64-bits wide, and cache lines (and memory bus) are as wide or wider.
Now they've extended that all the way to pointers and integers (and thus more ADDR lines for 64 (well 48-bit addressing), but that's a different story)
You can't do MMX and floating point in parallel. But you CAN do 64-bit arithmetic (which used to require MMX) and floating point in parallel, or 64-bit integer arithmetic and MMX/SSE/3dnow in parallel. This opens up the door for lots of additional speed optimizations by internal instruction level parallelism. Also it lets you do things with less instructions (and hence clocks) which used to be cumbersome before (faster fixed point/arbitrary precision math, large multiply/acc. operations, saturation math with 32-bit quantities)
... 64-bit bitfields!
Also another great one
but even when programming in assembler, it isn't so bad. In fact using the x86_64 IA assembler is pretty nice in long mode (AMD is happy to send you the manuals gratis, to help out). Not as symmetric as say MIPS, but it's getting better.
The IA is the IA. It has little to do with how anything else works (which is all modern at this point). The only pain is having non-fixed length opcodes, but we've solved that problem (TLBs) so we get to use instruction cache better. I don't really see the problem!