Storage speeds are irrelevant, you can add as much disk throughput as is necessary. Even pre-SSD, storage scales very very well. Its just not cheap, and it always has large access time.
"you've still only got a single north bridge": as you yourself say, wrong. You can attach a north bridge to each cpu, giving you a constant scaling factor of CPU's/external system I/O.
"Now we're back to coherency problems, but at two points.": Its allocating the work load on the NUMA system thats non-trivial. DragonFly BSD is the only reasonable & usable proposal I've seen that addresses this.
This is the poster child for vacuous argument on parallelism, and we'd be better if people would stop trying to say this.
The tasks you mention are laughably bad. You make up nonsense about new music codecs, pretend that indexing is something that happens all-the-time/realtime, and speak emptiness about "algorithms" that will magically consume cpu to make lives better.
The only viable thing you mention in your post is video editing, which is a task handled on the de-facto parallel processor of the computer: the GPU.
The notion that the average user is ever going to aggregate enough background work at the same time to need a multicore system is preposterous. Even this outlandish reference-free fantasy hasnt proposed a foreground task that a common end user would perform that would use even a mildly parallel cpu.
Niagra2 is 8 core, and 4 way SMT per core. Given that each core has multiple functional units, its very close to being a fully 32 way cpu. It feeds on four dual channel memory pipes. The servers running these dont need special software to make use of these cores, they just handle lots and lots and lots of user requests per second. For the most part they're webservers and fileservers, but they'd almost certainly make excellent mainframes for large multi user environments/virtualized systems. All it takes to use this multi core cpu is a dozen 5-watt 166mhz thin clients.
Besides multi user environments, theres already plenty of data parallel tasks that can use however many cores you give them. Graphics is just one, but neutral nets, signal processing, simulation tasks of all kinds, these tasks can all parallelized to any degree of parallelism you can build. 1000 will be laughably small by 2010; already today you can buy an 800 core cards in Best Buy: just ask for an AMD 4870 and pay your $200.
Yes applications today are mostly monolithic pieces of code running sequentially. Theres little we can do to speed these tasks up, its up to programmers to do a decent job of building reasonably responsive applications.
But the notion that code has to be parallelized to make us of multicores is patently false. Data parallel code turns a single computation into many threads of work, but all we really need is many threads of work to make use of a multicore system.
The real issue is that a single user might not be able to effectively utilize a multi core system. However, if you put thirty or a hundred users on a system, each users tasks aggregate to a sizable pool of threads that would be ideal for a multi core system.
The I/O bounds you mentioned are the real issue. Its largely the reason Nehalem is going with an integrated memory controller, to prevent the northbridge from limiting scaling. PPC, AMD, everyone else has gone this direction already, and for this reason.
Intel is switching to integrated memory controllers on each die with the upcoming Nehalem. It has a tri-channel memory controller that will eventually run 1.3GHz. Thats 32GB/s of memory bandwidth -per socket-. Get four Nehalems together and you'll have (almost) the same aggregate bandwidth as a reasonably top of the line graphics card.
Well, thats actually Linux's/lack of hardware documentation's fault. Almost all of these units have DMA engines onboard, but Linux will not use that DMA engine. If you want a finger to point at poor throughput, look to the kernel that has to do all data copying on the CPU.
when you buy a wireless router, just make sure its a router that will run a decent linux distribution. the linksys wrt54g started the ball rolling, and there is now a rather impressive list of routers supported by just one embedded linux distro; OpenWRT. dd-wrt has a similarly lengthy list. some allow you to attach hard drives via IDE or USB and do file serving as well. most run around 200mhz, have 4mb flash and 16/32mb ram, although better and worse configurations are available. these also have wireless built in, and usually two separate hardware vlans. you can pick up routers for under $50.
802.11n hardware seems to have very poor linux support, and not many routers have gigabit unfortunately. i havent really followed closely as neither of these features is on my "must have" list. the one i've seen moving recently is the wrt350n, which is making pretty good headway and has both features but its still not ready for primetime and is a pretty old router.
in general, i dont see why you'd get specific hardware for this when you could just have a small 5 watt linux router that handles your wan/lan/wifi/simple daemons.
| The problem is more along the lines of there are so many choices for centralized authentication in unix that there is no obvious default.
I beg to differ. LDAP+Kerberos is the way to do it. There really arent any network alternatives.
LDAP works fine, its setting up kerberos, samba, ftpd-of-choice, pam, your web server, vpn, dovecot, x509 certs, &c &c &c &C for ldap thats an enormous time consuming pain. LDAP itself gives you very little, its wiring up all the services that distros need to come along and provide for network users.
Theres a setting in Opera where you can tell it how much disk cache and how much memory cache to use. On my laptop with 232MB of ram I have Opera set to use 50MB of memory. Preferences -> Advanced, then, uh,, history? I think?
The fact that he only had 4 windows open is also pretty hilarious. I usually have a couple dozen windows open, and just opening Firefox 3 in that environment (not doing anything) takes about 900 MB of ram.
He makes no effort to find out whether "memory" being used is just buffered files (which can be dropped at any point) or data (which must be swapped out).
Also, Opera will use exactly as much memory as you tell it to. He makes no mention of how much memory he told Opera to use. I have a laptop with 232 MB of ram and I tell Opera to use 50MB there. It obliges.
It almost felt unfair working on my Computer Science degree with people who flat out hated computers and always wanted to copy each other's programming projects to pass classes, simply because they though that was the way to go for a good job. The industry could use a bit of thinning out if it means that we're left with actual bright and enthusiastic people who really do like doing this type of work.
I'm normally more worried about the people who dont hate computers. Most people with a genuine interest can rattle off a hateful list of grievances with modern software in their sleep. Just as much as it is the "dont care" people, it's the varieties of "shiny object" idolatry so many techies have that has kept computers from getting better in the last 10 years. Enthusiasm is not enough, we need critical eyes to re-evaluate how to do things better.
I dont have words to express how sorrowful it is that MS got integrated kerberos+ldap authentication+authorization before *nix world did. There really should be many distros with services that come out of box configured for kerberos+ldap. In casual conversation (elevator speeches) I used to call Linux a network OS, that it attempted to be a network citizen. The failure of Linux to converge on common network infrastructure is, to me, the most soul crushing failure I have witnessed from open source.
Dull unfortunately also means every company has to dedicate some % of the bottom line into a sunk cost called "IT" that they receive absolutely no boon from.
I'm not saying IT is useless or that there doesnt need to be some budget allocation, but the amount we spend maintaining and fixing infrastructure that should-- by now-- just work, is insane. Worse, I've seen very little improvement looking back 10 years.
This dull environment isnt good for business and its not good for IT. Obviously 10 different platforms by 10 different companies is not good or sustainable, but least common denominator stasis we've fallen towards isnt good for anyone either.
I dont believe the full on computer obsession is required to be an excellent computer using knowledge worker/IT agent. I think this is a common fallacy, and part of the reason people either think tech is the bee's knees or boring as crap.
The truth of the matter is that computers and the software on them are just tools. Further, our tools really havent changed that drastically in the last 10 years. Their interfaces may have, but for the most part the things you can do with computers are pretty similar to the things you could do with them 10 years ago.
So long as you have an active interest in learning to improve your tool using efficacy, I strongly believe it doesnt matter whether you are obsessed with learning about new tools or not. As long as you actively pursue means of making yourself and your environment better, I dont care whether you are writing in assembly, ruby, ocaml, or excel; you are the type of person who belongs in tech.
i think you'd have to use pretty mediocre quality settings to get that fps on a voodoo2. i went straight to geforce 1 sdr, and my dual 566mhz celeron bp6 still got slowdown on reasonable settings at 800x600.
you manage to miss every relevant point in the book in a very long elaboration of the status quo.
nvidia and amd and every consumer electronics company in the world are doing their damnedest to break that status quo and make your phone and everything else a capable all purpose platform. this nvidia chip can go in mobile phones, but its got a video engine capable of 1680x1050. why is that? because ~~***YOUR PHONE***~~ needs that display? good god no. the point is, we're seeing new embedded devices we expect to function in dual roles of a) phone and b) computer replacement.
long shaders let you do tasks like indirection in ways unfathomable for simpler setups. this in turn lets you run more application code in gpgpu land. this lets you save power. even if you disavow the use of it, i fail to understand how anyone could claim the lack of the feature is a good thing. it requires more advanced caching / buffering, but that should not be a dealbreaker. especially when we start loading our chips with massive onboard caches -- a secret well loved by the gamecube for example.
Close the lid. Does your computer last twice as long? No. I suggest before you shout your mouth off you attempt to gain some grounding in the topic. For starters, try:
watch -n1 cat/proc/acpi/battery/CMB1/state
With an absolute mizer of a cpu, a Crusoe 800, I go from 850ma to 550ma from the display. The reason the backlight is a 50% power hit is because the cpu is sucking almost nothing in the first place.
With LED backlighting or OLED the effect would be less pronounced.
The reason you dont notice movie cpu drain is because you are used to x86 cpus, which drain a lot of power whether you do anything with them or not. Thus some cycles spend on video decoding dont penalize your power consumption too much more. On a cpu like ARM, you would see very clear power differences between 60% cpu use v. 5% cpu use + hardware accel, as the CPU will basically be completely turned off 95% of the time (and all that time off will translate directly to power savings).
Dreamcast was Reneases with a PowerVR chip, iPhone is some bastard child Samsung ARM that i was told was doing CPU graphics... nope looks to be PowerVR. Ok, both PowerVR graphics, totally different cpu's though.
source code says "readable" source code will be posted soon.
i await that.
theres a lot of cool text interfaces happening on the web. theres in browser vi (jsvi), and source code editors like CodeMirror, CodePress, and more[1]. all very cool!
Slashdot Mirror
Storage speeds are irrelevant, you can add as much disk throughput as is necessary. Even pre-SSD, storage scales very very well. Its just not cheap, and it always has large access time.
"you've still only got a single north bridge": as you yourself say, wrong. You can attach a north bridge to each cpu, giving you a constant scaling factor of CPU's/external system I/O.
"Now we're back to coherency problems, but at two points.": Its allocating the work load on the NUMA system thats non-trivial. DragonFly BSD is the only reasonable & usable proposal I've seen that addresses this.
This is the poster child for vacuous argument on parallelism, and we'd be better if people would stop trying to say this.
The tasks you mention are laughably bad. You make up nonsense about new music codecs, pretend that indexing is something that happens all-the-time/realtime, and speak emptiness about "algorithms" that will magically consume cpu to make lives better.
The only viable thing you mention in your post is video editing, which is a task handled on the de-facto parallel processor of the computer: the GPU.
The notion that the average user is ever going to aggregate enough background work at the same time to need a multicore system is preposterous. Even this outlandish reference-free fantasy hasnt proposed a foreground task that a common end user would perform that would use even a mildly parallel cpu.
Niagra2 is 8 core, and 4 way SMT per core. Given that each core has multiple functional units, its very close to being a fully 32 way cpu. It feeds on four dual channel memory pipes. The servers running these dont need special software to make use of these cores, they just handle lots and lots and lots of user requests per second. For the most part they're webservers and fileservers, but they'd almost certainly make excellent mainframes for large multi user environments/virtualized systems. All it takes to use this multi core cpu is a dozen 5-watt 166mhz thin clients.
Besides multi user environments, theres already plenty of data parallel tasks that can use however many cores you give them. Graphics is just one, but neutral nets, signal processing, simulation tasks of all kinds, these tasks can all parallelized to any degree of parallelism you can build. 1000 will be laughably small by 2010; already today you can buy an 800 core cards in Best Buy: just ask for an AMD 4870 and pay your $200.
Yes applications today are mostly monolithic pieces of code running sequentially. Theres little we can do to speed these tasks up, its up to programmers to do a decent job of building reasonably responsive applications.
But the notion that code has to be parallelized to make us of multicores is patently false. Data parallel code turns a single computation into many threads of work, but all we really need is many threads of work to make use of a multicore system.
The real issue is that a single user might not be able to effectively utilize a multi core system. However, if you put thirty or a hundred users on a system, each users tasks aggregate to a sizable pool of threads that would be ideal for a multi core system.
The I/O bounds you mentioned are the real issue. Its largely the reason Nehalem is going with an integrated memory controller, to prevent the northbridge from limiting scaling. PPC, AMD, everyone else has gone this direction already, and for this reason.
Yes memory bandwidth is obviously a concern.
Intel is switching to integrated memory controllers on each die with the upcoming Nehalem. It has a tri-channel memory controller that will eventually run 1.3GHz. Thats 32GB/s of memory bandwidth -per socket-. Get four Nehalems together and you'll have (almost) the same aggregate bandwidth as a reasonably top of the line graphics card.
Well, thats actually Linux's/lack of hardware documentation's fault. Almost all of these units have DMA engines onboard, but Linux will not use that DMA engine. If you want a finger to point at poor throughput, look to the kernel that has to do all data copying on the CPU.
"Here;s an idea: Stop fucking stealing shit !! If you don't steal you won't care if your stealing facilitation enablers get a fucking RST or not. "
rst hurts anyone trying to keep long lived tcp connections, regardless of how much or what traffic they are sending.
why a device for just this?
when you buy a wireless router, just make sure its a router that will run a decent linux distribution. the linksys wrt54g started the ball rolling, and there is now a rather impressive list of routers supported by just one embedded linux distro; OpenWRT. dd-wrt has a similarly lengthy list. some allow you to attach hard drives via IDE or USB and do file serving as well. most run around 200mhz, have 4mb flash and 16/32mb ram, although better and worse configurations are available. these also have wireless built in, and usually two separate hardware vlans. you can pick up routers for under $50.
802.11n hardware seems to have very poor linux support, and not many routers have gigabit unfortunately. i havent really followed closely as neither of these features is on my "must have" list. the one i've seen moving recently is the wrt350n, which is making pretty good headway and has both features but its still not ready for primetime and is a pretty old router.
in general, i dont see why you'd get specific hardware for this when you could just have a small 5 watt linux router that handles your wan/lan/wifi/simple daemons.
wrong fucking topic my bad
wanna see something funny? try their hardware specs page. yeaaah we can definitely turn that into a great open source router.
Wanna see something funny?
hardware specs page.....
rriiiiiiggghhhhtttt clearly open source. we'll know exactly how to use this.
| The problem is more along the lines of there are so many choices for centralized authentication in unix that there is no obvious default.
I beg to differ. LDAP+Kerberos is the way to do it. There really arent any network alternatives.
LDAP works fine, its setting up kerberos, samba, ftpd-of-choice, pam, your web server, vpn, dovecot, x509 certs, &c &c &c &C for ldap thats an enormous time consuming pain. LDAP itself gives you very little, its wiring up all the services that distros need to come along and provide for network users.
Theres a setting in Opera where you can tell it how much disk cache and how much memory cache to use. On my laptop with 232MB of ram I have Opera set to use 50MB of memory. Preferences -> Advanced, then, uh,, history? I think?
The fact that he only had 4 windows open is also pretty hilarious. I usually have a couple dozen windows open, and just opening Firefox 3 in that environment (not doing anything) takes about 900 MB of ram.
He makes no effort to find out whether "memory" being used is just buffered files (which can be dropped at any point) or data (which must be swapped out).
Also, Opera will use exactly as much memory as you tell it to. He makes no mention of how much memory he told Opera to use. I have a laptop with 232 MB of ram and I tell Opera to use 50MB there. It obliges.
Article is pretty fucking crap ass.
It almost felt unfair working on my Computer Science degree with people who flat out hated computers and always wanted to copy each other's programming projects to pass classes, simply because they though that was the way to go for a good job. The industry could use a bit of thinning out if it means that we're left with actual bright and enthusiastic people who really do like doing this type of work.
I'm normally more worried about the people who dont hate computers. Most people with a genuine interest can rattle off a hateful list of grievances with modern software in their sleep. Just as much as it is the "dont care" people, it's the varieties of "shiny object" idolatry so many techies have that has kept computers from getting better in the last 10 years. Enthusiasm is not enough, we need critical eyes to re-evaluate how to do things better.I dont have words to express how sorrowful it is that MS got integrated kerberos+ldap authentication+authorization before *nix world did. There really should be many distros with services that come out of box configured for kerberos+ldap. In casual conversation (elevator speeches) I used to call Linux a network OS, that it attempted to be a network citizen. The failure of Linux to converge on common network infrastructure is, to me, the most soul crushing failure I have witnessed from open source.
Dull unfortunately also means every company has to dedicate some % of the bottom line into a sunk cost called "IT" that they receive absolutely no boon from.
I'm not saying IT is useless or that there doesnt need to be some budget allocation, but the amount we spend maintaining and fixing infrastructure that should-- by now-- just work, is insane. Worse, I've seen very little improvement looking back 10 years.
This dull environment isnt good for business and its not good for IT. Obviously 10 different platforms by 10 different companies is not good or sustainable, but least common denominator stasis we've fallen towards isnt good for anyone either.
I dont believe the full on computer obsession is required to be an excellent computer using knowledge worker/IT agent. I think this is a common fallacy, and part of the reason people either think tech is the bee's knees or boring as crap.
The truth of the matter is that computers and the software on them are just tools. Further, our tools really havent changed that drastically in the last 10 years. Their interfaces may have, but for the most part the things you can do with computers are pretty similar to the things you could do with them 10 years ago.
So long as you have an active interest in learning to improve your tool using efficacy, I strongly believe it doesnt matter whether you are obsessed with learning about new tools or not. As long as you actively pursue means of making yourself and your environment better, I dont care whether you are writing in assembly, ruby, ocaml, or excel; you are the type of person who belongs in tech.
I did not address tiler architecture in my comment. But yeah, I think we're near the same page here.
i think you'd have to use pretty mediocre quality settings to get that fps on a voodoo2. i went straight to geforce 1 sdr, and my dual 566mhz celeron bp6 still got slowdown on reasonable settings at 800x600.
you manage to miss every relevant point in the book in a very long elaboration of the status quo.
nvidia and amd and every consumer electronics company in the world are doing their damnedest to break that status quo and make your phone and everything else a capable all purpose platform. this nvidia chip can go in mobile phones, but its got a video engine capable of 1680x1050. why is that? because ~~***YOUR PHONE***~~ needs that display? good god no. the point is, we're seeing new embedded devices we expect to function in dual roles of a) phone and b) computer replacement.
long shaders let you do tasks like indirection in ways unfathomable for simpler setups. this in turn lets you run more application code in gpgpu land. this lets you save power. even if you disavow the use of it, i fail to understand how anyone could claim the lack of the feature is a good thing. it requires more advanced caching / buffering, but that should not be a dealbreaker. especially when we start loading our chips with massive onboard caches -- a secret well loved by the gamecube for example.
A large amount of .NET programs will "just run" on Windows CE.
No.
/proc/acpi/battery/CMB1/state
Close the lid. Does your computer last twice as long? No. I suggest before you shout your mouth off you attempt to gain some grounding in the topic. For starters, try:
watch -n1 cat
With an absolute mizer of a cpu, a Crusoe 800, I go from 850ma to 550ma from the display. The reason the backlight is a 50% power hit is because the cpu is sucking almost nothing in the first place.
With LED backlighting or OLED the effect would be less pronounced.
The reason you dont notice movie cpu drain is because you are used to x86 cpus, which drain a lot of power whether you do anything with them or not. Thus some cycles spend on video decoding dont penalize your power consumption too much more. On a cpu like ARM, you would see very clear power differences between 60% cpu use v. 5% cpu use + hardware accel, as the CPU will basically be completely turned off 95% of the time (and all that time off will translate directly to power savings).
Dreamcast was Reneases with a PowerVR chip, iPhone is some bastard child Samsung ARM that i was told was doing CPU graphics... nope looks to be PowerVR. Ok, both PowerVR graphics, totally different cpu's though.
source code says "readable" source code will be posted soon.
i await that.
theres a lot of cool text interfaces happening on the web. theres in browser vi (jsvi), and source code editors like CodeMirror, CodePress, and more[1]. all very cool!
[1] http://en.wikipedia.org/wiki/Comparison_of_Javascript-based_source_code_editors