I'm sure there's some company out there with some kind of HDD robot so that they don't have to spend people down to switch hard drives in their giant storage arrays.
But why would they switch hard drives? Just let them sit there broken and start new ones up. Once a month or a year or whatever you can send someone in to replace the broken ones, that should not take long.
Rumour has it that Google never replaces broken hardware, it just sits there dead until Google upgrades the whole rack.
It'll cost less than disk drives and still be reasonably durable.
That's the problem. Tapes are at best half the price of disks, and the drive is expensive. Next year disks have fallen in price, but the tapes still cost the same, and that just gets worse over time. To maintain a decent tape price it is necessary to switch standard quite often.
HTC phones are unreliable and need charging hourly. I say that as an HTC owner. We're seeing Nokia vs. Ericsson all over, with Samsung as Nokia and HTC as Ericsson. Back then I ended up choosing Ericsson again and again for the cool features, every time believing that the new model would not have the same problems.
I probably won't make the same mistake with HTC though. They have shown that they have no interest in providing firmware upgrades, even though they promised to do so. My next phone will be Android (but not HTC), and if that one doesn't get firmware updates I'll be switching away from Android.
At what point does the bandwidth/throughput of the DMA start limiting the performance of your backup?
With USB2? Never, it does PIO. The CPU gets to babysit the transfer. That is very much the exception in the PC world of today, every other bus in use today offloads nicely.
In general, storage buses on PC's are rarely fast enough to saturate system buses, and therefore DMA does not limit throughput. If you are really lucky you can get perhaps 2GBps through a fast PC, but you can probably copy memory around at 10 times that speed. So just like the mainframe, it will be limited by the performance of the slowest devices multiplied by parallelism.
With two tars going on different devices, either the reading tar or the writing tar will be slowest, and the other will not contribute anything much to the wall clock time. Compare to cp -r which requires threading (or built-in multiple processes, or async I/O) to achieve the same speed. I bet GNU cp isn't using any of those.
I would expect the built-in pipe buffer in a modern OS to be large enough on its own, but perhaps mbuffer will help. On Linux you can adjust the pipe buffer size with fcntl, so you could implement the equivalent of mbuffer with just a small preloaded library.
He's doing the same number of source opens and destinations creates, but he's adding a step.
An extremely cheap step, free in fact because the task is I/O limited. If cp is unable to read one file while writing the next (quite likely), it will lose badly.
The Pi only decode H.263 and H.264 in hardware. (Well this is not entirely correct, but you need a license to unlock the other formats, and there is no way to get the license...)
I'll take an endless cycle of 'universe lifetimes', thank you. Modern astronomy is hugely depressing -- no, we didn't get a nice cyclical universe, no, we didn't get a slow "heat death" when entropy goes to infinity. Instead, everything just flies to bits, so far away from everything else that it is impossible to even see the other galaxies while the last stars burn out. Even if there is still intelligent life somewhere at that point, they have no way to figure out that anything but the local galaxy ever existed (if we're lucky and dark energy doesn't overpower that too).
And if they do somehow know what came before, if someone kept records, will that not be even more depressing?
Read googles hard disk failure analysis. Smart is a poor indicator of disk failure. Most failed disks in my experience do give no warning
Exactly!
According to Googles data, you can expect SMART data to change in less than half of the cases where drives die, and less than half of the cases where SMART warns actually lead to drive failures.
My Intel X25-M died on me a few weeks ago. It started doing write errors, then some reads started failing too. It was possible to read all but two files off the/home partition. This was with the original firmware, I never got around to doing the performance upgrade.
It is nasty to have a drive die, but there are certainly worse failures than that one.
When inventing a new programming language, you just need to write a reasonable interpreter or a compiler to show the world how fantastic your new language is. You can do that pretty quickly, and it is a mandatory part of most computer science degrees.
When inventing a new database language, you either have to start by spending a long time figuring out how an existing database backend works, or you need to write a new one from scratch. Most people give up before they get to write the first line of actual language code.
In addition, few people actually use the power of SQL. Most just use it as a stupid data dump and handle consistency etc. in application code.
Otherwise the banks and other financial institutions would be repudiating stock market trades made by their automated systems but which they later wished had not been made.
This happens quite regularly. Not every day, because it is still quite labour intensive and problematic for the exchanges to roll back transactions, but it does happen.
The guy sitting at the Tokyo exchange does not yet know that US company X has gone bankrupt. If you see value of a stock you hold falling dramatically in New York, you want to sell the same stock in Tokyo before anyone there notices. If you are lucky you can unload enough stock to save you from losing a fortune. Similarly for sudden gains which have not yet moved to the other exchanges.
I'm probably setting myself up for a *whoosh* here, but you can't do that. When you push one end of a rod, you are actually creating a wave in the material. No material can be stiff enough that the wave travels faster than c. Come to think of it, maybe I'm setting myself up for a joke here, rather than a *whoosh*.
I'm sorry but your argument makes no sense. Imagine a communication line consisting of 100km fiber from the New York exchange, neutrino link from New York to Tokyo, 100km fiber to the Tokyo exchange. With sufficiently fast neutrino generators and detectors, that means a latency in the 30ms range compared to an ideal latency in the 70ms range using conventional fiber.
If you suddenly discover that your shares in company X are worthless, it is great to have 40ms to get rid of them before anyone else knows.
Osborning isn't about changing strategy. That's a perfectly reasonable response to realising you have a bad strategy.
If Nokia switched to Android now, they would be admitting that their upcoming Windows 8-based products aren't good enough, but that their new Android-based phones are going to be fantastic. That is precisely Osborning. Of the options for avoiding Osborning you propose, only supporting the old systems is realistic.
Anyway, it is academic. Nokia cannot afford to go without the money injections from Microsoft, unless they give up on producing phones and turn into a pure patent-troll.
Nokia still has around 30,000 patents left. At 750kUSD each that comes to 22.5GUSD. They apparently have an income of 0.5GUSD/year from the patents right now which is a bit on the low side if they are as valuable as the Motorola or Nortel patents. However, that does not include the money saved by not having to license those patents from someone else while making mobile phones.
At 10GUSD market cap, Nokia is considerably cheaper than Motorola and Nortel were. Especially since Nokia is still sitting on a 5GUSD pile of cash. Buying Nokia and laying off everyone except the lawyers sounds like an excellent deal right now.
Slashdot Mirror
I'm sure there's some company out there with some kind of HDD robot so that they don't have to spend people down to switch hard drives in their giant storage arrays.
But why would they switch hard drives? Just let them sit there broken and start new ones up. Once a month or a year or whatever you can send someone in to replace the broken ones, that should not take long.
Rumour has it that Google never replaces broken hardware, it just sits there dead until Google upgrades the whole rack.
It'll cost less than disk drives and still be reasonably durable.
That's the problem. Tapes are at best half the price of disks, and the drive is expensive. Next year disks have fallen in price, but the tapes still cost the same, and that just gets worse over time. To maintain a decent tape price it is necessary to switch standard quite often.
The last point in particular is why you don't see HDD robots: all that handling would skyrocket the hard drives' failure rate.
No, you don't see HDD robots because the "reader" is approximately free. There is no advantage to moving the drive and disconnecting the cabling.
Modern laptop drives are at least as robust as tapes when turned off, you could easily make a hard drive robot.
HTC phones are unreliable and need charging hourly. I say that as an HTC owner. We're seeing Nokia vs. Ericsson all over, with Samsung as Nokia and HTC as Ericsson. Back then I ended up choosing Ericsson again and again for the cool features, every time believing that the new model would not have the same problems.
I probably won't make the same mistake with HTC though. They have shown that they have no interest in providing firmware upgrades, even though they promised to do so. My next phone will be Android (but not HTC), and if that one doesn't get firmware updates I'll be switching away from Android.
GCC is GPL too, yet many commercial products use it.
The little library GCC uses for builtin functions etc. is not under the GPL.
At what point does the bandwidth/throughput of the DMA start limiting the performance of your backup?
With USB2? Never, it does PIO. The CPU gets to babysit the transfer. That is very much the exception in the PC world of today, every other bus in use today offloads nicely.
In general, storage buses on PC's are rarely fast enough to saturate system buses, and therefore DMA does not limit throughput. If you are really lucky you can get perhaps 2GBps through a fast PC, but you can probably copy memory around at 10 times that speed. So just like the mainframe, it will be limited by the performance of the slowest devices multiplied by parallelism.
With two tars going on different devices, either the reading tar or the writing tar will be slowest, and the other will not contribute anything much to the wall clock time. Compare to cp -r which requires threading (or built-in multiple processes, or async I/O) to achieve the same speed. I bet GNU cp isn't using any of those.
I would expect the built-in pipe buffer in a modern OS to be large enough on its own, but perhaps mbuffer will help. On Linux you can adjust the pipe buffer size with fcntl, so you could implement the equivalent of mbuffer with just a small preloaded library.
He's doing the same number of source opens and destinations creates, but he's adding a step.
An extremely cheap step, free in fact because the task is I/O limited. If cp is unable to read one file while writing the next (quite likely), it will lose badly.
Taliban had almost eradicated opium production before the US invaded. That made them rather unpopular with the farmers.
In 2017 we will hear "5 years of falling temperatures" and in 2022 "no temperature rise for 10 years" and so on. Just like 1998.
How to move the cursor one character forward in vi
The Pi only decode H.263 and H.264 in hardware. (Well this is not entirely correct, but you need a license to unlock the other formats, and there is no way to get the license...)
Some people are trying to decode MPEG-2 in software.
If only it could play 480i video without a problem...
I'll take an endless cycle of 'universe lifetimes', thank you. Modern astronomy is hugely depressing -- no, we didn't get a nice cyclical universe, no, we didn't get a slow "heat death" when entropy goes to infinity. Instead, everything just flies to bits, so far away from everything else that it is impossible to even see the other galaxies while the last stars burn out. Even if there is still intelligent life somewhere at that point, they have no way to figure out that anything but the local galaxy ever existed (if we're lucky and dark energy doesn't overpower that too).
And if they do somehow know what came before, if someone kept records, will that not be even more depressing?
Read googles hard disk failure analysis. Smart is a poor indicator of disk failure. Most failed disks in my experience do give no warning
Exactly!
According to Googles data, you can expect SMART data to change in less than half of the cases where drives die, and less than half of the cases where SMART warns actually lead to drive failures.
My Intel X25-M died on me a few weeks ago. It started doing write errors, then some reads started failing too. It was possible to read all but two files off the /home partition. This was with the original firmware, I never got around to doing the performance upgrade.
It is nasty to have a drive die, but there are certainly worse failures than that one.
When inventing a new programming language, you just need to write a reasonable interpreter or a compiler to show the world how fantastic your new language is. You can do that pretty quickly, and it is a mandatory part of most computer science degrees.
When inventing a new database language, you either have to start by spending a long time figuring out how an existing database backend works, or you need to write a new one from scratch. Most people give up before they get to write the first line of actual language code.
In addition, few people actually use the power of SQL. Most just use it as a stupid data dump and handle consistency etc. in application code.
Otherwise the banks and other financial institutions would be repudiating stock market trades made by their automated systems but which they later wished had not been made.
This happens quite regularly. Not every day, because it is still quite labour intensive and problematic for the exchanges to roll back transactions, but it does happen.
The guy sitting at the Tokyo exchange does not yet know that US company X has gone bankrupt. If you see value of a stock you hold falling dramatically in New York, you want to sell the same stock in Tokyo before anyone there notices. If you are lucky you can unload enough stock to save you from losing a fortune. Similarly for sudden gains which have not yet moved to the other exchanges.
I'm probably setting myself up for a *whoosh* here, but you can't do that. When you push one end of a rod, you are actually creating a wave in the material. No material can be stiff enough that the wave travels faster than c. Come to think of it, maybe I'm setting myself up for a joke here, rather than a *whoosh*.
I'm sorry but your argument makes no sense. Imagine a communication line consisting of 100km fiber from the New York exchange, neutrino link from New York to Tokyo, 100km fiber to the Tokyo exchange. With sufficiently fast neutrino generators and detectors, that means a latency in the 30ms range compared to an ideal latency in the 70ms range using conventional fiber.
If you suddenly discover that your shares in company X are worthless, it is great to have 40ms to get rid of them before anyone else knows.
Osborning isn't about changing strategy. That's a perfectly reasonable response to realising you have a bad strategy.
If Nokia switched to Android now, they would be admitting that their upcoming Windows 8-based products aren't good enough, but that their new Android-based phones are going to be fantastic. That is precisely Osborning. Of the options for avoiding Osborning you propose, only supporting the old systems is realistic.
Anyway, it is academic. Nokia cannot afford to go without the money injections from Microsoft, unless they give up on producing phones and turn into a pure patent-troll.
No, I don't have a good source. There is a page, Why Are Launch Costs So High" which looks at the problem.
I got it from an article named Nokia prepared to sell patents if price right: CFO. That refers to the first quarter 2012 report. I found this line in the Nokia q1 2012 interim report: "We estimate that our current annual IPR royalty income run-rate is approximately EUR 0.5 billion."
Nokia still has around 30,000 patents left. At 750kUSD each that comes to 22.5GUSD. They apparently have an income of 0.5GUSD/year from the patents right now which is a bit on the low side if they are as valuable as the Motorola or Nortel patents. However, that does not include the money saved by not having to license those patents from someone else while making mobile phones.
At 10GUSD market cap, Nokia is considerably cheaper than Motorola and Nortel were. Especially since Nokia is still sitting on a 5GUSD pile of cash. Buying Nokia and laying off everyone except the lawyers sounds like an excellent deal right now.