Re:Most programs don't need a 64-bit address space
on
Linux 3.4 Released
·
· Score: 2
You've seen the prices of 16+ GB of ram recently, right?
16GB is pretty cheap, beyond that it starts to get expensive because you either need either expensive 8GB modules or a high end CPU platform (intel LGA2011 or amd G34) with more ram channels.
If you plan for all your ram to be used by one big process thaen x32 won't be of interest to you. OTOH if you are serving webapps (lots of processes but each individual process not using much ram) written in languages that make heavy use of pointers then x32 starts to look attractive.
not in your data, which is far and away larger
All depends on what form your data is in, if it's just one big lump of processor-independent data then you probablly won't gain much from x32.
OTOH if it's complex structures with lots of pointers then using a system with 32-bit pointers will reduce it's size considerablly.
Re:btrfs needed the work
on
Linux 3.4 Released
·
· Score: 4, Informative
This is known as featuritis, and is anathema to the Unix way, where each part should do just one thing, and do it extremely well.
The problem with conventional raid is it has no way of knowing which redundant copy of the information is correct and indeed it may well end up overwriting the correct copy with the bad copy during a resync. So it protects against drives that fail but it doesn't protect against drives that quietly return bad data.
In theory you could implement a raid layer with strong checksums so it knew which copy was bad, but the problem then becomes where to put those checksums (without creating a load of extra seeks).
By implementing raid techniques as part of the filesystem the checksums can be stored with the existing metadata. Implementing raid as part of the filesystem also allows different redundancy policies to be applied to different data.
just look at the R-Pi running 1080p H.264 video over HDMI
Afiact the only reason the Pi can play 1080p H.264 acceptablly it is because it's decoded by the videocore GPU. The arm is nowhere near powerful enough.
Which is all well and good if all you want to play is H.264 but if you want to play anything else you are at the mercy of the device vendor (the Pi foundation have talked about selling an additional codec pack for the videocore but it's unclear whether it will actually happen) and if you want to do something other than 3D graphics or playing video then the videocore can't help you at all.
Also while "ram is cheap" for intel/amd systems that doesn't seem to be the case for arm systems. Nearly every arm system i've looked into had it's ram soldered to the board and few have more than half a gigabyte.
The first thing I always recommend to anyone getting unsatisfactory speeds is rewire your telephone sockets and place the modem as close to the master socket as possible.
and put all the telephone wiring in your house behind a single decent quality filter rather than using seperate filters for each phone. Branches are BAD for high frequency signals.
Having said that a massive performance drop at particular times of day sounds more like a congestion issue than a line issue to me.
I think the honest answer is that while there may be some techincal pros/cons with DSL vs cable what really matters is how shitty the provider supplying the service is. Regardless of whether the service is cable or DSL if the provider cares to provide a decent service they can. If the provider doesn't care to provide a decent service then you likely won't get one
The equation you are looking for P=(V^2)CF which is derived from combining the equations P=IV and I=VCF and provides a reqasonable approximatino for how digital CMOS power consumption will behave.
But it's only an approximation for a couple of reasons
1: I=VCF gives an average current. P=IV is true for instantanious voltage and instantanious current It is only correct for average current if votlage is a constant. Assuming that voltage is constant is an approximation of reality. 2: I=VCF assumes that all the current gpes to charging and discharging capacitance. In relality some also goes to leakage and some may also go to shootthrough especially at higher voltages.
Capacitance is static
Not really. The total capacitance may be fixed but what matters for the equation is the ammount charged on each cycle and that can change heavilly with what the chip is doing.
A chip may draw amperage, but that is just a function of C and F.
No, current is a vital step in understanding the limitations of the equation you are using.
P.S. maybe it's an american thing (but i'm a brit) but to me using the term "amperage" rather than current makes you sound like you don't have a clue.
AIUI the problem is you can't tell the difference externally between a weak but unshielded source and between a strong source in a thick lead box.
So if you want to find the people who are moving the strong sources around you have to use a sensitive detector and then investigate to find out exactly what is being carried.
How do we know in 60 years time the JPEG standard will still be around?
We don't know for sure but given how widely supported it is I think that barring the complete collapse of civilisation it is highly unlikely that we will lose the ability to decode it.
This isn't about getting "my own way." This is about keeping the country from turning into fucking greece.
There is a big difference between the US and greece.
The US has a lot of debt but it's denominated in US dollars which the US congress has the power to control the "printing" of. The greek governement has a lot of debt and it is denominated in euros which they do not have the power to control the printing of.
If you have BOTH mirroring/reconstructive parity and things are set up properly so they work together*
oops, didn't complete that bit, it should have said
If you have BOTH checksumming and mirroring/reconstructive parity and things are set up properly so they work together* then you can detect which block is bad and then take action to recover the data.
The real trick is to combine checksuming with mirroring or reconstructive parity.
If you only have mirroring/reconstructive parity you can reconstruct the data if you know which block is bad If you only have checksumming you can detect which block is bad but you can't do anything about it. If you have BOTH mirroring/reconstructive parity and things are set up properly so they work together*
Afaict this was/is the killer feature of zfs, I think btrfs now has it too but i'm not positive (though the impression i've got is that btrfs is still too unstable generally to be used as a main fs).
*Having a raid system below a checksumming system doesn't work for this because the raid system may overwrite the good data with the bad during a resync. Having a checksumming system below a raid system would work but runs into the problem that checksumming is difficult to do efficiently below the filesystem layer (because you need to store the checksums somewhere) so the only real way to do it efficiently and correctly is to integrate both checksumming and redundancy into the filesystem.
identification puts much higher requirements on a hash algorithm than corruption detection.
Lets connsider corruption detection.
Say you have c corruputed files. Now suppose our hash has m possibilities all equally likely (the ideal case for a hash function). The average number of files that will pass the hash test despite being corrupt is c/m
Now lets consider identification
Say you have n different files, the number of possible pairs of files is ((n^2)-n)/2 . For large n we can approximate this as n^2/2 . Now suppose our hash has m possibilities all equally likely (the ideal case for a hash function). The probability of any pair files having colliding hashes colliding is 1/m So the average number of collisions is approximately (n^2)/(2h).
Do the sums and you find you need a much larger hash to get acceptable performance in an identification application than in a corruption detection application.
Note: all the above assumes no malice is involved. If malice is involved then the requirements on the hash get much tighter.
IMO when considering renting something, buying something on credit or getting an upfront discount for locking yoursel into a contract you should be asking yourself "how much will it really improve my financial situation and/or quality of life to get this now rather that when I can afford to buy it outright", "is that improvement worth the extra cost" and "is my financial situation stable enough to take on the commitment".
P.S. I think that any advert for a service with a minimum contract term should be required to include the total ammount payable in their adverts just like adverts for purchaes on credit do (at least in the UK).
The biggest flaw with the N9 was that the OS was NOT a major OS
The thing is I don't see WP7 as a major OS either. MS were late to the "multitouch+decent browser" smartphone market and then threw away the goodwill they had when they threw out all support for applications from thier pre-mulitouch smartphone platforn and replaced it with a locked down.net environment.
The older nokia phones had user replacable covers that were a good mm or so from the screen. So you could crack the cover (say by sitting on your coat with your phone and keys in the pocket) and the screen would still be fine. with the newer phones that is not the case (mine is on it's third screen)
Oh and with the old phones you could disable the backlight and the screen was still perfectly usable, again can't do that with the modern ones.
Mmm, for an existing provider the question is not "how much are we making now?", it's "how much more would we make if we offered faster services and/or wider coverage and how much would it cost us to do that?"
And for a new upstart the question is "can we displace enough customers from the incumbent to pay for our fixed costs?"
How much does going from say 5mbps to 50mbps improve the internet experiance for normal users? how much extra do you think most people would pay for that improvement?
Telecom has used 23" racks for years. There are standard adapters already available to mount 19" hardware in a 23" rack. 23" racks are already available in the marketplace.
AIUI they don't want to increase the overall size of the rack so they can fit into standard datacenter floorplans, presumablly this limits the ammount they can widen the server space in the rack without widening the rack itself.
Telecom has been using DC plant for decades and there is a ton of existing 48Vdc equipment on the market.
48V DC was well suited to the needs of telcos at the time who originally used it to directly power POTs lines, that doesn't mean it's appropriate for powering computers.
To prevent currents flowing through metalwork causing ground potential shifts and to limit damage in the event of faults you generally want your computing gear to be isolated from your building power distribution system. This generally implies having an isolating PSU (whether AC-DC or DC-DC) between your building power distribution and your motherboards.
Afaict this 12V system is about connecting multiple servers in a rack to a common rack level PSU, not about changing building power distribution standards. Presumablly 12V was chosen beause it is becoming the standard for powering motherboards (5V and 3.3V are mostly legacy supplies nowadays)
hmm, I have an older macbook with intel graphics and i've had quite a bit of trouble with xorg on it. I've found a version that mostly works and i'm scared to upgrade further for fear of breaking it. Maybe i'm just unlucky.
Not that ATI and NVIDIA are great either, back when I last worked with an ATI card in a linux box FGLRX had a nasty habbit of crashing the system when it couldn't find it's module and the nvidia linux binary driver wouldn't work with my latest card (an EVGA GT 430 dual DVI, interestingly the EVGA GT 430 DVI+VGA my brother has works fine) meaning I had to fall back to using nouveau (I play my 3D games under windows so it's not critical but it certainly irks me).
Having been burnt by all three vendors i'm not sure what i'll do for my next machine.
But a current gen celeron or pentium can hardly be overclocked at all because intel has locked the multipliers and tied the clocks of timing sensitive interfaces to BCLK (making any significant overclock by BCLK impossible).
So rather than an overclock taking you from "low end" to "pretty high end" it takes you from "pretty high end" to "faster in threadcount limited tasks than any CPU running at it's stock speed".
Apple takes techology designed and manufactured elsewhere (though they may make some minor tweaks), combines it with their own software and turns it into slick products that people are prepared to pay a premium for. That requires design and marketing prowess but not any great technicalcapability.
You do realise that width is not the only charactersitic of a rack right??
In particular rackmount servers tend to be really deep compared to pretty much every other rackmount device, so if you are going to use them in a rack that wasn't specifically designed for servers you need to be really carefull about depth.
This probablly varies with where you live. Here in the UK you tend to need a phone line for in with most internet packages (or at least not having a phone line doesn't save you much) and there are many cheap calling packages that only cover landline numbers. Further landline call quality is better than mobile so may as well use the landline.
Regardless I agree cat5e/6 everywhere is the way to go for distributing communication arround the house. Still you want to make sure you have appropriate connections (or prefferablly a duct you can pull stuff through from wherever communication services are likely to enter your house to whereever your main rack is.
No it didn't at least not at the cheap end of the market, most new monitors now have DVI but I see plenty of older ones (both CRT and LCD) that only have VGA and are still perfectly servicable.
Or do you also want your DB9 serial, DB25 parallel and PS/2 ports back?
my latest motherboard (gigabyte Z68AP-D3) has one each of those ports and I wasn't even specifically looking for any of them when I bought it. It doesn't have VGA but there are other motherboards in the same price range that do (I don't use the onboard graphics so I didn't care what interface it had).
Compatibility with pretty much every monitor in circulation. There are lots of older/cheaper flat panels in circulation that don't have a digital input and there are still a few old CRTs hanging arronund as well.
Having to buy a monitor VS being able to use a castoff monitor makes a big difference to the total cost.
Slashdot Mirror
You've seen the prices of 16+ GB of ram recently, right?
16GB is pretty cheap, beyond that it starts to get expensive because you either need either expensive 8GB modules or a high end CPU platform (intel LGA2011 or amd G34) with more ram channels.
If you plan for all your ram to be used by one big process thaen x32 won't be of interest to you. OTOH if you are serving webapps (lots of processes but each individual process not using much ram) written in languages that make heavy use of pointers then x32 starts to look attractive.
not in your data, which is far and away larger
All depends on what form your data is in, if it's just one big lump of processor-independent data then you probablly won't gain much from x32.
OTOH if it's complex structures with lots of pointers then using a system with 32-bit pointers will reduce it's size considerablly.
This is known as featuritis, and is anathema to the Unix way, where each part should do just one thing, and do it extremely well.
The problem with conventional raid is it has no way of knowing which redundant copy of the information is correct and indeed it may well end up overwriting the correct copy with the bad copy during a resync. So it protects against drives that fail but it doesn't protect against drives that quietly return bad data.
In theory you could implement a raid layer with strong checksums so it knew which copy was bad, but the problem then becomes where to put those checksums (without creating a load of extra seeks).
By implementing raid techniques as part of the filesystem the checksums can be stored with the existing metadata. Implementing raid as part of the filesystem also allows different redundancy policies to be applied to different data.
just look at the R-Pi running 1080p H.264 video over HDMI
Afiact the only reason the Pi can play 1080p H.264 acceptablly it is because it's decoded by the videocore GPU. The arm is nowhere near powerful enough.
Which is all well and good if all you want to play is H.264 but if you want to play anything else you are at the mercy of the device vendor (the Pi foundation have talked about selling an additional codec pack for the videocore but it's unclear whether it will actually happen) and if you want to do something other than 3D graphics or playing video then the videocore can't help you at all.
Also while "ram is cheap" for intel/amd systems that doesn't seem to be the case for arm systems. Nearly every arm system i've looked into had it's ram soldered to the board and few have more than half a gigabyte.
The first thing I always recommend to anyone getting unsatisfactory speeds is rewire your telephone sockets and place the modem as close to the master socket as possible.
and put all the telephone wiring in your house behind a single decent quality filter rather than using seperate filters for each phone. Branches are BAD for high frequency signals.
Having said that a massive performance drop at particular times of day sounds more like a congestion issue than a line issue to me.
I think the honest answer is that while there may be some techincal pros/cons with DSL vs cable what really matters is how shitty the provider supplying the service is. Regardless of whether the service is cable or DSL if the provider cares to provide a decent service they can. If the provider doesn't care to provide a decent service then you likely won't get one
P=VC^2F
That equation is plain wrong.
The equation you are looking for P=(V^2)CF which is derived from combining the equations P=IV and I=VCF and provides a reqasonable approximatino for how digital CMOS power consumption will behave.
But it's only an approximation for a couple of reasons
1: I=VCF gives an average current. P=IV is true for instantanious voltage and instantanious current It is only correct for average current if votlage is a constant. Assuming that voltage is constant is an approximation of reality.
2: I=VCF assumes that all the current gpes to charging and discharging capacitance. In relality some also goes to leakage and some may also go to shootthrough especially at higher voltages.
Capacitance is static
Not really. The total capacitance may be fixed but what matters for the equation is the ammount charged on each cycle and that can change heavilly with what the chip is doing.
A chip may draw amperage, but that is just a function of C and F.
No, current is a vital step in understanding the limitations of the equation you are using.
P.S. maybe it's an american thing (but i'm a brit) but to me using the term "amperage" rather than current makes you sound like you don't have a clue.
AIUI the problem is you can't tell the difference externally between a weak but unshielded source and between a strong source in a thick lead box.
So if you want to find the people who are moving the strong sources around you have to use a sensitive detector and then investigate to find out exactly what is being carried.
How do we know in 60 years time the JPEG standard will still be around?
We don't know for sure but given how widely supported it is I think that barring the complete collapse of civilisation it is highly unlikely that we will lose the ability to decode it.
This isn't about getting "my own way." This is about keeping the country from turning into fucking greece.
There is a big difference between the US and greece.
The US has a lot of debt but it's denominated in US dollars which the US congress has the power to control the "printing" of. The greek governement has a lot of debt and it is denominated in euros which they do not have the power to control the printing of.
If you have BOTH mirroring/reconstructive parity and things are set up properly so they work together*
oops, didn't complete that bit, it should have said
If you have BOTH checksumming and mirroring/reconstructive parity and things are set up properly so they work together* then you can detect which block is bad and then take action to recover the data.
The real trick is to combine checksuming with mirroring or reconstructive parity.
If you only have mirroring/reconstructive parity you can reconstruct the data if you know which block is bad
If you only have checksumming you can detect which block is bad but you can't do anything about it.
If you have BOTH mirroring/reconstructive parity and things are set up properly so they work together*
Afaict this was/is the killer feature of zfs, I think btrfs now has it too but i'm not positive (though the impression i've got is that btrfs is still too unstable generally to be used as a main fs).
*Having a raid system below a checksumming system doesn't work for this because the raid system may overwrite the good data with the bad during a resync. Having a checksumming system below a raid system would work but runs into the problem that checksumming is difficult to do efficiently below the filesystem layer (because you need to store the checksums somewhere) so the only real way to do it efficiently and correctly is to integrate both checksumming and redundancy into the filesystem.
identification puts much higher requirements on a hash algorithm than corruption detection.
Lets connsider corruption detection.
Say you have c corruputed files.
Now suppose our hash has m possibilities all equally likely (the ideal case for a hash function).
The average number of files that will pass the hash test despite being corrupt is c/m
Now lets consider identification
Say you have n different files, the number of possible pairs of files is ((n^2)-n)/2 . For large n we can approximate this as n^2/2 .
Now suppose our hash has m possibilities all equally likely (the ideal case for a hash function). The probability of any pair files having colliding hashes colliding is 1/m
So the average number of collisions is approximately (n^2)/(2h).
Do the sums and you find you need a much larger hash to get acceptable performance in an identification application than in a corruption detection application.
Note: all the above assumes no malice is involved. If malice is involved then the requirements on the hash get much tighter.
IMO when considering renting something, buying something on credit or getting an upfront discount for locking yoursel into a contract you should be asking yourself "how much will it really improve my financial situation and/or quality of life to get this now rather that when I can afford to buy it outright", "is that improvement worth the extra cost" and "is my financial situation stable enough to take on the commitment".
P.S. I think that any advert for a service with a minimum contract term should be required to include the total ammount payable in their adverts just like adverts for purchaes on credit do (at least in the UK).
The biggest flaw with the N9 was that the OS was NOT a major OS
The thing is I don't see WP7 as a major OS either. MS were late to the "multitouch+decent browser" smartphone market and then threw away the goodwill they had when they threw out all support for applications from thier pre-mulitouch smartphone platforn and replaced it with a locked down .net environment.
pity the newer ones aren't.
The older nokia phones had user replacable covers that were a good mm or so from the screen. So you could crack the cover (say by sitting on your coat with your phone and keys in the pocket) and the screen would still be fine. with the newer phones that is not the case (mine is on it's third screen)
Oh and with the old phones you could disable the backlight and the screen was still perfectly usable, again can't do that with the modern ones.
Mmm, for an existing provider the question is not "how much are we making now?", it's "how much more would we make if we offered faster services and/or wider coverage and how much would it cost us to do that?"
And for a new upstart the question is "can we displace enough customers from the incumbent to pay for our fixed costs?"
How much does going from say 5mbps to 50mbps improve the internet experiance for normal users? how much extra do you think most people would pay for that improvement?
Telecom has used 23" racks for years. There are standard adapters already available to mount 19" hardware in a 23" rack. 23" racks are already available in the marketplace.
AIUI they don't want to increase the overall size of the rack so they can fit into standard datacenter floorplans, presumablly this limits the ammount they can widen the server space in the rack without widening the rack itself.
Telecom has been using DC plant for decades and there is a ton of existing 48Vdc equipment on the market.
48V DC was well suited to the needs of telcos at the time who originally used it to directly power POTs lines, that doesn't mean it's appropriate for powering computers.
To prevent currents flowing through metalwork causing ground potential shifts and to limit damage in the event of faults you generally want your computing gear to be isolated from your building power distribution system. This generally implies having an isolating PSU (whether AC-DC or DC-DC) between your building power distribution and your motherboards.
Afaict this 12V system is about connecting multiple servers in a rack to a common rack level PSU, not about changing building power distribution standards. Presumablly 12V was chosen beause it is becoming the standard for powering motherboards (5V and 3.3V are mostly legacy supplies nowadays)
hmm, I have an older macbook with intel graphics and i've had quite a bit of trouble with xorg on it. I've found a version that mostly works and i'm scared to upgrade further for fear of breaking it. Maybe i'm just unlucky.
Not that ATI and NVIDIA are great either, back when I last worked with an ATI card in a linux box FGLRX had a nasty habbit of crashing the system when it couldn't find it's module and the nvidia linux binary driver wouldn't work with my latest card (an EVGA GT 430 dual DVI, interestingly the EVGA GT 430 DVI+VGA my brother has works fine) meaning I had to fall back to using nouveau (I play my 3D games under windows so it's not critical but it certainly irks me).
Having been burnt by all three vendors i'm not sure what i'll do for my next machine.
But a current gen celeron or pentium can hardly be overclocked at all because intel has locked the multipliers and tied the clocks of timing sensitive interfaces to BCLK (making any significant overclock by BCLK impossible).
So rather than an overclock taking you from "low end" to "pretty high end" it takes you from "pretty high end" to "faster in threadcount limited tasks than any CPU running at it's stock speed".
Apple takes techology designed and manufactured elsewhere (though they may make some minor tweaks), combines it with their own software and turns it into slick products that people are prepared to pay a premium for. That requires design and marketing prowess but not any great technicalcapability.
You do realise that width is not the only charactersitic of a rack right??
In particular rackmount servers tend to be really deep compared to pretty much every other rackmount device, so if you are going to use them in a rack that wasn't specifically designed for servers you need to be really carefull about depth.
This probablly varies with where you live. Here in the UK you tend to need a phone line for in with most internet packages (or at least not having a phone line doesn't save you much) and there are many cheap calling packages that only cover landline numbers. Further landline call quality is better than mobile so may as well use the landline.
Regardless I agree cat5e/6 everywhere is the way to go for distributing communication arround the house. Still you want to make sure you have appropriate connections (or prefferablly a duct you can pull stuff through from wherever communication services are likely to enter your house to whereever your main rack is.
VGA = that shit died a decade ago.
No it didn't at least not at the cheap end of the market, most new monitors now have DVI but I see plenty of older ones (both CRT and LCD) that only have VGA and are still perfectly servicable.
Or do you also want your DB9 serial, DB25 parallel and PS/2 ports back?
my latest motherboard (gigabyte Z68AP-D3) has one each of those ports and I wasn't even specifically looking for any of them when I bought it. It doesn't have VGA but there are other motherboards in the same price range that do (I don't use the onboard graphics so I didn't care what interface it had).
Compatibility with pretty much every monitor in circulation. There are lots of older/cheaper flat panels in circulation that don't have a digital input and there are still a few old CRTs hanging arronund as well.
Having to buy a monitor VS being able to use a castoff monitor makes a big difference to the total cost.
I don't think the debian amd64 port was released until 2007 ish, I think as part of Debian 4.0/etch
There was a semi-official amd64 port of sarge but etch was indeed the first release where it was included in the official archive.