In what way? You have yet to quantify this or the "ton of space" assertion. I'm not saying that density can't matter, just that it doesn't in the practical case. The only time I've heard otherwise is from vendors (hardware or datacenter) pushing a high-density product (which was never any cheaper than the less-dense alternative).
> And 12 individual disk will have a high failure rate than one disk because there is 12x electronics, 12x motors, 12x seals. Also the power needs for 12 disks is higher than 1. And you would need more raid cpu power to coordinate 12x disks.
Of course, but these are all calculable costs (although CPU cost for concatenation is de minimis). So, again, do you just not even perform that calculation?
> I feel like you are just arguing for arguing sake here.
I'm just trying to clarify what point you're making when you specify a particular solution, especially a RAID, which carries an associated cost, in the context of a dicussion about the economy of various options.
> If it was significantly cheaper and new sure. Ddr3 is just as expensive as ddr4 right now.
I maintain that new doesn't matter (see below). As for significantly cheaper, if it's recycled (be it already purchased but unused, spares, or working pulls), it would be a form of gratis-free. If a vendor has significant stock of now-undesirable lower-density or lower-speed memory modules (maybe even bare chips), that can result in a steep discount. It doesn't really matter if it's DDR3 or DDR4.
> No body running a professional data center wants the unpredictability of used parts unless they have had very consistent and tame histories. According to large scale studies ddr has increasing error rates as they age: https://ai.google/research/pubs/pub35162
Thanks for that study. I was unaware memory suffered from aging quite that much. That's further reason for disqualifying from recycling parts of unknown provenance (since they could be pulls due to a UE), but such parts would be disqualified anyway
Still, I fear that you're forgetting we're talking about swap here, not main memory. There's much more wiggle room for doing something like hot-sparing, even without otherwise having a RAID, since the performance of the backend storage (RAM) dwarfs the performance of the frontend (PCIe).
> Conclusion 3: The incidence of CEs increases with age, while the incidence of UEs decreases with age (due to re- placements).
This conclusion also suggests that there's a skew toward early mortality, so a pool of used, known-good DIMMs would actually be more reliable (and predictable, based on this study) than a pool of brand new ones.
I think we can agree to disagree here for the most part.
It was an interesting discussion. I do think there is a market for another type of volitile storage in PCs as long as DDR prices stay sky high as they are now, faster than SSD and slower than DDR4, and much cheaper than DDR4, close to the prices of SSD and with much higher read-write capabilities.
The problem with SSDs is their limited read-write count I guess as their cost are actually sort of excellent these days.
With regards to hard drives I actually follow backblaze's lead, thus it is received knowledge. They tend to buy the largest drives they can that are not premium priced:
Well, the market certainly agrees with you..there has never, to my knowledge, been anything like what I describe, other than for trivial capacities.
> With regards to hard drives I actually follow backblaze's lead, thus it is received knowledge. They tend to buy the largest drives they can that are not premium priced
This is completely valid for comparable use cases (which I think are vanishingly rare.. usually, performance matters). Their data is, of course, invaluable to the tech community at large (and potentially an annoyance to drive manufacturers).
In what way? You have yet to quantify this or the "ton of space" assertion. I'm not saying that density can't matter, just that it doesn't in the practical case. The only time I've heard otherwise is from vendors (hardware or datacenter) pushing a high-density product (which was never any cheaper than the less-dense alternative).
> And 12 individual disk will have a high failure rate than one disk because there is 12x electronics, 12x motors, 12x seals. Also the power needs for 12 disks is higher than 1. And you would need more raid cpu power to coordinate 12x disks.
Of course, but these are all calculable costs (although CPU cost for concatenation is de minimis). So, again, do you just not even perform that calculation?
> I feel like you are just arguing for arguing sake here.
I'm just trying to clarify what point you're making when you specify a particular solution, especially a RAID, which carries an associated cost, in the context of a dicussion about the economy of various options.
> If it was significantly cheaper and new sure. Ddr3 is just as expensive as ddr4 right now.
I maintain that new doesn't matter (see below). As for significantly cheaper, if it's recycled (be it already purchased but unused, spares, or working pulls), it would be a form of gratis-free. If a vendor has significant stock of now-undesirable lower-density or lower-speed memory modules (maybe even bare chips), that can result in a steep discount. It doesn't really matter if it's DDR3 or DDR4.
> No body running a professional data center wants the unpredictability of used parts unless they have had very consistent and tame histories. According to large scale studies ddr has increasing error rates as they age: https://ai.google/research/pubs/pub35162
Thanks for that study. I was unaware memory suffered from aging quite that much. That's further reason for disqualifying from recycling parts of unknown provenance (since they could be pulls due to a UE), but such parts would be disqualified anyway
Still, I fear that you're forgetting we're talking about swap here, not main memory. There's much more wiggle room for doing something like hot-sparing, even without otherwise having a RAID, since the performance of the backend storage (RAM) dwarfs the performance of the frontend (PCIe).
> Conclusion 3: The incidence of CEs increases with age, while the incidence of UEs decreases with age (due to re- placements).
This conclusion also suggests that there's a skew toward early mortality, so a pool of used, known-good DIMMs would actually be more reliable (and predictable, based on this study) than a pool of brand new ones.