3 ms·

SSDs have been mainstream products for at least a decade now. Ten years ago, the largest HDD we had was like 2 TB. We currently have 16 TB HDDs and people are e

by Hamuko 6y ago

SSDs have been mainstream products for at least a decade now. Ten years ago, the largest HDD we had was like 2 TB. We currently have 16 TB HDDs and people are expecting us to have 50 TB HDDs in 2026. And not only that, 16 TB HDD is quite affordable. Just 500 dollars for a Seagate IronWolf. Meanwhile a 3.2 TB Seagate Nytro will set you back about $1800. SSD mass storage is an option for only those with deep pockets. Hell, not even tape drives have gone away yet. Amazon Glacier is just one example.

pg-gadfly 6y ago
It's also to note that SSD capacity hasn't strictly improved, it's been gained through massive compromises that severely impact their lifetime and performance. More than that, a 5 year old nas drive can outperform a modern cheap ssd in throughput easily.
- wtallis 6y ago
  Most, but not all of the capacity and affordability increases of SSDs have been due to compromises elsewhere. If you plot the characteristics of "the cheapest flash you can buy" over the years, few of the curves are monotonic. Physical memory cell dimensions decreased steadily until the arrival of 3D NAND, whereupon cell sizes jumped back up and have been relatively stable since. SSDs incorporate more advanced error correction with each generation of controllers, so today's drives can get more usable write endurance out of the same physical cell size. Number of bits that can be stored per cell has been increasing. Each time we cram another bit into each memory cell, performance and write endurance drop. But in between those major shifts, successive generations of flash tend to get faster and more energy-efficient over time. Today's 3 bit per cell TLC NAND is much faster than the TLC from 5 years ago, and has better write endurance thanks to the transition to 3D NAND. Today's SLC NAND has comparable performance and write endurance to the planar SLC that disappeared from the market over 5 years ago, but now it's several times cheaper. And most of the compromises that have been made with NAND flash over the past 10+ years are purely academic. The vast majority of use cases do not need flash rated for 100k P/E cycles, and excess endurance beyond your needs provides no tangible benefit. Most of the endurance that has theoretically been sacrificed was never needed in the first place. Most of the performance lost has been offset by using more flash in parallel and by eliminating bottlenecks elsewhere. The only metric by which a 5 year old NAS hard drive—or any hard drive—will outperform a modern cheap SSD is in long-term sustained sequential write throughput. Even the slowest modern cheap SSD will offer at least twice the sequential read throughput, and orders of magnitude better random read or write throughput. And it's not like the write throughput of modern SSDs is abysmal: gigabit Ethernet is a more significant bottleneck when filling a 1TB or larger dirt-cheap consumer SSD. So unless you're trying to use cheap consumer SSDs in an expensive NAS with 10GbE, the poor write performance of low-end SSDs won't come into play.
  - pg-gadfly 6y ago
    Of course there has been improvement in absolute terms, but our current state is much more than just that, which skews the results when thinking about where we are. The opinion is definitely not popular on hackernews, but I still want to throw it out there. Cheap SSDs can tank below 70MB/s after cache, which is less than a gigabit connection, so it can totally affect your performance even in the smallest local networks. You don't have to sustain writes for long at all to start seeing the compromises hidden in cheap SSDs, especially if they're using their TLC as SLC cache, in which case you run out faster and faster.
eropple 6y ago
I was under the impression that Glacier is hard drive backed, not tape backed?
- jasonwatkinspdx 6y ago
  Yeah, it's spinning disk backed. What's different is the drives are aggregated up on hosts under the design assumption that most of the drives will be powered off most of the time, allowing a lower cost per byte.
- Hamuko 6y ago
  Some industry insider has claimed that it's tape-backed. There's also been speculation that it's actually based on optical media.
  - londons_explore 6y ago
    It's totally backed on "we have loads of hard drives with spare storage capacity but no IO capacity, so sell it off cheap, but make sure it doesn't perform too well or it'll eat into our other products". Nearly all cloud providers have more storage bytes than they have IO (ie. each individual hard drive receives so many read/write requests per second that if they filled the drive all the way to full some read/write requests couldn't be handled and would be unacceptably delayed)

by Hamuko 6y ago

pg-gadfly 6y ago

It's also to note that SSD capacity hasn't strictly improved, it's been gained through massive compromises that severely impact their lifetime and performance. More than that, a 5 year old nas drive can outperform a modern cheap ssd in throughput easily.

wtallis 6y ago
Most, but not all of the capacity and affordability increases of SSDs have been due to compromises elsewhere. If you plot the characteristics of "the cheapest flash you can buy" over the years, few of the curves are monotonic. Physical memory cell dimensions decreased steadily until the arrival of 3D NAND, whereupon cell sizes jumped back up and have been relatively stable since. SSDs incorporate more advanced error correction with each generation of controllers, so today's drives can get more usable write endurance out of the same physical cell size. Number of bits that can be stored per cell has been increasing. Each time we cram another bit into each memory cell, performance and write endurance drop. But in between those major shifts, successive generations of flash tend to get faster and more energy-efficient over time. Today's 3 bit per cell TLC NAND is much faster than the TLC from 5 years ago, and has better write endurance thanks to the transition to 3D NAND. Today's SLC NAND has comparable performance and write endurance to the planar SLC that disappeared from the market over 5 years ago, but now it's several times cheaper. And most of the compromises that have been made with NAND flash over the past 10+ years are purely academic. The vast majority of use cases do not need flash rated for 100k P/E cycles, and excess endurance beyond your needs provides no tangible benefit. Most of the endurance that has theoretically been sacrificed was never needed in the first place. Most of the performance lost has been offset by using more flash in parallel and by eliminating bottlenecks elsewhere. The only metric by which a 5 year old NAS hard drive—or any hard drive—will outperform a modern cheap SSD is in long-term sustained sequential write throughput. Even the slowest modern cheap SSD will offer at least twice the sequential read throughput, and orders of magnitude better random read or write throughput. And it's not like the write throughput of modern SSDs is abysmal: gigabit Ethernet is a more significant bottleneck when filling a 1TB or larger dirt-cheap consumer SSD. So unless you're trying to use cheap consumer SSDs in an expensive NAS with 10GbE, the poor write performance of low-end SSDs won't come into play.
- pg-gadfly 6y ago
  Of course there has been improvement in absolute terms, but our current state is much more than just that, which skews the results when thinking about where we are. The opinion is definitely not popular on hackernews, but I still want to throw it out there. Cheap SSDs can tank below 70MB/s after cache, which is less than a gigabit connection, so it can totally affect your performance even in the smallest local networks. You don't have to sustain writes for long at all to start seeing the compromises hidden in cheap SSDs, especially if they're using their TLC as SLC cache, in which case you run out faster and faster.

eropple 6y ago

I was under the impression that Glacier is hard drive backed, not tape backed?

jasonwatkinspdx 6y ago
Yeah, it's spinning disk backed. What's different is the drives are aggregated up on hosts under the design assumption that most of the drives will be powered off most of the time, allowing a lower cost per byte.
Hamuko 6y ago
Some industry insider has claimed that it's tape-backed. There's also been speculation that it's actually based on optical media.
- londons_explore 6y ago
  It's totally backed on "we have loads of hard drives with spare storage capacity but no IO capacity, so sell it off cheap, but make sure it doesn't perform too well or it'll eat into our other products". Nearly all cloud providers have more storage bytes than they have IO (ie. each individual hard drive receives so many read/write requests per second that if they filled the drive all the way to full some read/write requests couldn't be handled and would be unacceptably delayed)