I'm not sure why you felt the need to mention QEMU specifically.
I suspect because the article incorrectly describes it as an emulator, while it is capable of full virtualization if the plugin the GP post linked to is used.
It's not the only such mistake in the article: Xen is described as performing paravirtualization, but it too is capable of full virtualization in some cases (i.e., when it is supported by the hardware).
As a spare-time kernel writer, I couldn't agree more. The problem is that a few small improvements like this won't make a new kernel viable. It needs to be a radical improvement over existing systems. Witness Plan9.
CPUs put a lot of stock in branch prediction; due to the nature of OOP languages like C++, Objective-C (I like this one), and D, this doesn't work.
CPU designers have been working on this problem for a while now. Most modern CPUs can branch-predict C++-style code with a fairly high degree of accuracy. Prediction misses generally only occur when the class of an object being processed is different to the class of the last object to be processed. This case would likely cause a prediction miss whatever approach was taken to implement it.
I've been noticing significantly longer start-up times for Java over the years.
The main driver for this is increasing complexity of the class library. The java.lang.Object class now depends directly or indirectly on over 200 other classes. These classes must all be loaded and initialized before any user code can run.
Most modern processors come with a vector unit that can handle vectors of 32-bit integers, so if you have to handle large quantities of text you can speed certain things up even more by running streaming calculation on the vector unit.
Yes, but you could speed it up more by using vectors of 8 or 16 bit units.
Today's computers have more than 2,000 times as much memory as the machines of yesteryear, yet programmers are still writing code as if memory is in short supply.
I stopped reading after that first line.
Yeah, it was pretty dumb. I mean, we've been building digital computers for over 50 years now. You'd have thought by now that we'd have written all the software we'd ever want, right?
No. We keep writing new software because we're doing new things that older computers couldn't do. That means that all that memory is necessary, because otherwise the application we're working on would have been written 20 years ago.
IMHO, CS graduates these days do not understand the implications of the code they are writing.
IMHO, lots of programmers today don't understand the power of the tools they use.
For example... I did a code review a few months ago where in one place that was inside a FOR Loop, there was a lot of statements that always returned the same values and never used any of the values that changed between iterations of the loop.
Did you try compiling it and looking at the resulting code? Most modern compilers automatically perform an optimization called loop-invariant code motion, which takes that stuff that does the same calculation on each iteration and only runs it once at the start.
Harvard architecture chips would still be susceptible to "return-to-libc" style exploit code: overwrite the return address with the address of code already on the system that you want to execute, and follow it with the parameters you want to execute it with.
[0] well, unless it's a PIC microcontroller with a separate stack memory.
I understand that a number of more recent CPU architectures have taken the step of putting the return address into a register, rather than on the stack. Itanium is an example.
GC is an obvious requirement of a "safe" language.
No, it isn't. Memory safety can be achieved without the use of garbage collection, by avoid reallocation of freed memory to a differently-typed object. Access to freed pointers can be caught by page table manipulation. Sure, there's an overhead to these techniques, but then there's a non-trivial overhead to GC as well.
I'm sure that somebody is going to dig up that paging-free GC paper, but pay attention: that is a kernel-level GC.
Which just indicates that there's been a lack of forethought in OS design resulting in the impossibility of implementing this at user level. I'm not sure, but pretty confident it could be done in user mode in a well-designed microkernel (e.g. L4). Not sure how relevant this is at this point, though. We've gone too far down the current line of OS development to just drop everything in favour of a new design just to be able to do a couple of neat tricks like this.
Complaint 3: Swing is ugly/leaks memory
The first is a matter of opinion.
Yes, but it's a very popular opinion. Swing (and JDK) is also exceptionally slow. Most people don't realise it, but almost all of the slowness traditionally associated with Java can be attributed to those toolkits. Don't believe me? Try writing a program with another toolkit. I haven't tried it myself, but you could give wxjava a go.
Complaint 4: Bad build system
Java cannot do incremental builds if class files have circular references. In a small project of about ten classes I was working on, the only way to build it was "rm *.class ; javac *.java"
Have you tried jikes? It has a superior incemental build system to sun's compiler. Not sure it'll fix your issue, but anything's possible.
What programs are you running? If it's a typical setup, with some browser with a few tabs, a common viruscan (norton home versions), aim, maybe messenger, WMP/Winamp/Itunes, limewire, then you'll notice usage going up.
That particularly machine is only used for batch processing: it's currently got a single browser window (IE6), bash shell and video encoding processes running.
I'm using Windows Vista Ultimate and im using 573 megs right now. IE 7 with several tabs open right now, (haven't closed it in a few days) is running 167 megs of ram.
And people call Firefox a memory hog. That's using 80 megs on my main desktop right now, with 7 tabs open and has been running uninterrupted for 3 days, during which it has seen substantial use.
Out of curiosity, when was the last time you were using a machine with 1gig of ram?
My main desktop has the most RAM of any machine I've ever had sole use of, and that's 416MB (256+128+32). It rarely swaps, despite running most of the types of application you mention. Not the same apps: notably I use utorrent not limewire, and don't have antivirus running as a service, both of which provide noticeable savings. My commit charge hasn't gone above my physical limit since last reboot, according to task manager. I currently have the best part of 90MB available.
However, speedy, is an entirely different story. Swapping takes a lot of time, compare to using real ram, not virtual ram.
The machine's total RAM usage ("commit charge" in MS-speak) is 145MB. Over the last 5 minutes of usage, including starting up the task manager to find that usage figure, plus the extra overhead of the performance monitoring system, it has averaged 4 page faults per second, with a peak of 100. The performance is pretty good.
"Will Windows 2000 have the drivers running at ring-0?"
The answer was initially "what does that mean?" and eventually, "yes it will."
This isn't Win32's only source of vulnerability and instability, but it's certainly among them.
Running non-priveleged drivers is a troublesome prospect. It isn't technically hard. MS could certainly do it. The issue is performance. It means performing an extra context switch on each transition into/out of a driver. Context switches are notoriously slow operations. User-mode device drivers therefore perform slowly, unless the system is particularly well optimized. L4 is about the only system to achieve it. Mach, which OSX is based on comes close, but is generally regarded as having failed.
And when MacOSX came out and did precisely what I had hoped Windows would do [...]
OSX runs device drivers in priveleged mode, just the same as Windows does. See The diagram on this page.
Regarding having a "backwards compatibility mode" - the Windows kernel has built-in support for this kind of thing. It's used to run Win16 programs, OS/2 programs and (some) POSIX-style programs. MS could use it in a future release to add an entirely new API, if they wished. I don't suspect it would be hard at all, and there'd be no backwards-compatibility issues to deal with.
Eveyone will switch to 64bit hardware by the end of 2008, it's impossible otherwise. The moore's law tells us what will be the memory capacity by 2008. And with 32bit it's impossible to address more than 4GB.
Moores law tells us about availability of memory; it doesn't tell us anything about what applications will drive demand, and without demand nobody will shell out cash for extra RAM. My guess is that takeup will slow. Filling more than 512MB requires demanding (or badly written) applications.
Also, current-gen 32-bit machines can address 16GB, adding another 3 years until we hit maximum, even at full Moore's Law rate.
Investors care about returns. Growth is only one way for that to happen. A steady stream of large dividends from fairly constant profits will keep them happy too.
Those people would have paid $300 or so per copy, which is money out of M$'s pocket
I doubt many of these people will pay $300 for a copy of Vista. Some of them, for instance, seem to be mac users who would never do it. Others will choose not to upgrade until they buy a new machine, so will get an OEM copy, for which MS will likely only see ~$100. Others still will skip vista and not upgrade until the next version is released, which is unlikely to take as long as vista did.
Flourescent is more efficient than semiconductor lighting.
Depends. Fluorescent is slightly more efficent than pure-white LEDs (which are based on an absorbtion-reemission system). But for coloured lighting, LEDs are substantially superior. Looks as though LEDs are going to start winning on the white scale in the next few years, too. You can now get 64 lumens/watt (compare with 60 typical from CF and 80 from traditional fluorescent) from a white LED, and the figures keep looking better. LEDs are also instant-on and have an even-longer lifetime than CF.
It would be nice if power sockets here in N. America came fitted with individual on/off switches, as is standard in the UK - these devices could continue to live plugged in to the wall, but with their consumption more easily controlled.
Unfortunately the design of most power adapters means you can't reach the switch while they're plugged in anyway.
Apart from usually being a ridiculous single-piece design that occludes several other sockets in a power strip, they cause massive cable tangles and practical use requires that they be left permanently powered-on.
Checking mine with a power usage meter suggests somewhere around 3/4 of them are high enough quality that they don't draw power when not loaded up.
Slashdot Mirror
I'm not sure why you felt the need to mention QEMU specifically.
I suspect because the article incorrectly describes it as an emulator, while it is capable of full virtualization if the plugin the GP post linked to is used.
It's not the only such mistake in the article: Xen is described as performing paravirtualization, but it too is capable of full virtualization in some cases (i.e., when it is supported by the hardware).
As a spare-time kernel writer, I couldn't agree more. The problem is that a few small improvements like this won't make a new kernel viable. It needs to be a radical improvement over existing systems. Witness Plan9.
They should complain to their network provider - it's where the problem is.
Tell you what, you get a major ISP to fix a broken mail server with only their customer services phone number to start from.
Basically, the users get shafted. RBLs are a bad idea because of this.
CPUs put a lot of stock in branch prediction; due to the nature of OOP languages like C++, Objective-C (I like this one), and D, this doesn't work.
CPU designers have been working on this problem for a while now. Most modern CPUs can branch-predict C++-style code with a fairly high degree of accuracy. Prediction misses generally only occur when the class of an object being processed is different to the class of the last object to be processed. This case would likely cause a prediction miss whatever approach was taken to implement it.
I've been noticing significantly longer start-up times for Java over the years.
The main driver for this is increasing complexity of the class library. The java.lang.Object class now depends directly or indirectly on over 200 other classes. These classes must all be loaded and initialized before any user code can run.
Most modern processors come with a vector unit that can handle vectors of 32-bit integers, so if you have to handle large quantities of text you can speed certain things up even more by running streaming calculation on the vector unit.
Yes, but you could speed it up more by using vectors of 8 or 16 bit units.
D itself conflicts with a pre-existing language design: a replacement for SQL designed by Hugh Darwen and Chris Date.
metaphors, you have to understand them.
Judging by that article, I think a more appropriate response would be to wipe them off the face of the Earth.
Today's computers have more than 2,000 times as much memory as the machines of yesteryear, yet programmers are still writing code as if memory is in short supply.
I stopped reading after that first line.
Yeah, it was pretty dumb. I mean, we've been building digital computers for over 50 years now. You'd have thought by now that we'd have written all the software we'd ever want, right?
No. We keep writing new software because we're doing new things that older computers couldn't do. That means that all that memory is necessary, because otherwise the application we're working on would have been written 20 years ago.
IMHO, CS graduates these days do not understand the implications of the code they are writing.
IMHO, lots of programmers today don't understand the power of the tools they use.
For example...
I did a code review a few months ago where in one place that was inside a FOR Loop, there was a lot of statements that always returned the same values and never used any of the values that changed between iterations of the loop.
Did you try compiling it and looking at the resulting code? Most modern compilers automatically perform an optimization called loop-invariant code motion, which takes that stuff that does the same calculation on each iteration and only runs it once at the start.
Harvard architecture chips would still be susceptible to "return-to-libc" style exploit code: overwrite the return address with the address of code already on the system that you want to execute, and follow it with the parameters you want to execute it with.
[0] well, unless it's a PIC microcontroller with a separate stack memory.
I understand that a number of more recent CPU architectures have taken the step of putting the return address into a register, rather than on the stack. Itanium is an example.
GC is an obvious requirement of a "safe" language.
No, it isn't. Memory safety can be achieved without the use of garbage collection, by avoid reallocation of freed memory to a differently-typed object. Access to freed pointers can be caught by page table manipulation. Sure, there's an overhead to these techniques, but then there's a non-trivial overhead to GC as well.
I'm sure that somebody is going to dig up that paging-free GC paper, but pay attention: that is a kernel-level GC.
Which just indicates that there's been a lack of forethought in OS design resulting in the impossibility of implementing this at user level. I'm not sure, but pretty confident it could be done in user mode in a well-designed microkernel (e.g. L4). Not sure how relevant this is at this point, though. We've gone too far down the current line of OS development to just drop everything in favour of a new design just to be able to do a couple of neat tricks like this.
Complaint 3: Swing is ugly/leaks memory
The first is a matter of opinion.
Yes, but it's a very popular opinion. Swing (and JDK) is also exceptionally slow. Most people don't realise it, but almost all of the slowness traditionally associated with Java can be attributed to those toolkits. Don't believe me? Try writing a program with another toolkit. I haven't tried it myself, but you could give wxjava a go.
Complaint 4: Bad build system
Java cannot do incremental builds if class files have circular references. In a small project of about ten classes I was working on, the only way to build it was "rm *.class ; javac *.java"
Have you tried jikes? It has a superior incemental build system to sun's compiler. Not sure it'll fix your issue, but anything's possible.
What programs are you running? If it's a typical setup, with some browser with a few tabs, a common viruscan (norton home versions), aim, maybe messenger, WMP/Winamp/Itunes, limewire, then you'll notice usage going up.
That particularly machine is only used for batch processing: it's currently got a single browser window (IE6), bash shell and video encoding processes running.
I'm using Windows Vista Ultimate and im using 573 megs right now. IE 7 with several tabs open right now, (haven't closed it in a few days) is running 167 megs of ram.
And people call Firefox a memory hog. That's using 80 megs on my main desktop right now, with 7 tabs open and has been running uninterrupted for 3 days, during which it has seen substantial use.
Out of curiosity, when was the last time you were using a machine with 1gig of ram?
My main desktop has the most RAM of any machine I've ever had sole use of, and that's 416MB (256+128+32). It rarely swaps, despite running most of the types of application you mention. Not the same apps: notably I use utorrent not limewire, and don't have antivirus running as a service, both of which provide noticeable savings. My commit charge hasn't gone above my physical limit since last reboot, according to task manager. I currently have the best part of 90MB available.
However, speedy, is an entirely different story. Swapping takes a lot of time, compare to using real ram, not virtual ram.
The machine's total RAM usage ("commit charge" in MS-speak) is 145MB. Over the last 5 minutes of usage, including starting up the task manager to find that usage figure, plus the extra overhead of the performance monitoring system, it has averaged 4 page faults per second, with a peak of 100. The performance is pretty good.
I raised my hand to ask a simple question:
"Will Windows 2000 have the drivers running at ring-0?"
The answer was initially "what does that mean?" and eventually, "yes it will."
This isn't Win32's only source of vulnerability and instability, but it's certainly among them.
Running non-priveleged drivers is a troublesome prospect. It isn't technically hard. MS could certainly do it. The issue is performance. It means performing an extra context switch on each transition into/out of a driver. Context switches are notoriously slow operations. User-mode device drivers therefore perform slowly, unless the system is particularly well optimized. L4 is about the only system to achieve it. Mach, which OSX is based on comes close, but is generally regarded as having failed.
And when MacOSX came out and did precisely what I had hoped Windows would do [...]
OSX runs device drivers in priveleged mode, just the same as Windows does. See The diagram on this page.
Regarding having a "backwards compatibility mode" - the Windows kernel has built-in support for this kind of thing. It's used to run Win16 programs, OS/2 programs and (some) POSIX-style programs. MS could use it in a future release to add an entirely new API, if they wished. I don't suspect it would be hard at all, and there'd be no backwards-compatibility issues to deal with.
Eveyone will switch to 64bit hardware by the end of 2008, it's impossible otherwise. The moore's law tells us what will be the memory capacity by 2008. And with 32bit it's impossible to address more than 4GB.
Moores law tells us about availability of memory; it doesn't tell us anything about what applications will drive demand, and without demand nobody will shell out cash for extra RAM. My guess is that takeup will slow. Filling more than 512MB requires demanding (or badly written) applications.
Also, current-gen 32-bit machines can address 16GB, adding another 3 years until we hit maximum, even at full Moore's Law rate.
Windows XP dies with only 256 megs of ram, its constantly swapping, and this is only after opening one or two additonal apps
I run XP on a machine with 128MB, running an app that typically consumes around 90. Everything's fine and stable.
No, snort doesn't work on XP since MS disabled raw sockets.
Oh, wait...
Investors care about growth.
Investors care about returns. Growth is only one way for that to happen. A steady stream of large dividends from fairly constant profits will keep them happy too.
Those people would have paid $300 or so per copy, which is money out of M$'s pocket
I doubt many of these people will pay $300 for a copy of Vista. Some of them, for instance, seem to be mac users who would never do it. Others will choose not to upgrade until they buy a new machine, so will get an OEM copy, for which MS will likely only see ~$100. Others still will skip vista and not upgrade until the next version is released, which is unlikely to take as long as vista did.
Flourescent is more efficient than semiconductor lighting.
Depends. Fluorescent is slightly more efficent than pure-white LEDs (which are based on an absorbtion-reemission system). But for coloured lighting, LEDs are substantially superior. Looks as though LEDs are going to start winning on the white scale in the next few years, too. You can now get 64 lumens/watt (compare with 60 typical from CF and 80 from traditional fluorescent) from a white LED, and the figures keep looking better. LEDs are also instant-on and have an even-longer lifetime than CF.
It would be nice if power sockets here in N. America came fitted with individual on/off switches, as is standard in the UK - these devices could continue to live plugged in to the wall, but with their consumption more easily controlled.
Unfortunately the design of most power adapters means you can't reach the switch while they're plugged in anyway.
Apart from usually being a ridiculous single-piece design that occludes several other sockets in a power strip, they cause massive cable tangles and practical use requires that they be left permanently powered-on.
Checking mine with a power usage meter suggests somewhere around 3/4 of them are high enough quality that they don't draw power when not loaded up.
When you mix with creme or milk, the coffee comes down to a very reasonable temperature.
Would it really hurt for a member of staff to do this, rather than leaving it to the customer?