User: Password:
|
|
Subscribe / Log in / New account

The Tux3 filesystem returns

The Tux3 filesystem returns

Posted Jan 1, 2013 19:04 UTC (Tue) by juliank (subscriber, #45896)
Parent article: The Tux3 filesystem returns

If everything is delayed, will this make it possible to add support to the file system that allows it to commit data only every 5 minutes or something like this -- that would allow the disk to be powered down in between, and thus reduce power usage on laptops.

I don't know whether any file system supports something like this already, but I know that ext4 doesn't do it (even if you set commit=<time> and prevent fsync and stuff, it will still commit before <time> has passed).


(Log in to post comments)

The Tux3 filesystem returns

Posted Jan 1, 2013 21:01 UTC (Tue) by daniel (guest, #3181) [Link]

Tux3 takes complete control of when delta transitions take place. After a delta transition (end of one delta, beginning of another) the "marshal" step sets up and queues a delta commit. No delta transition, no marshal, no delta commit, no disk writes, simple. I'm sure other filesystems are capable of this level of control, but in Tux3 the entire commit pipeline is controlled by this one thing.

The Tux3 filesystem returns

Posted Jan 1, 2013 21:58 UTC (Tue) by Richard_J_Neill (subscriber, #23093) [Link]

Would 5-minute commits still make sense on SSD?
By the time this hits mainstream, will there be any laptops with rotating storage?

The Tux3 filesystem returns

Posted Jan 1, 2013 23:36 UTC (Tue) by dgm (subscriber, #49227) [Link]

Millions. Seriously.

Displacing HDs

Posted Jan 2, 2013 0:19 UTC (Wed) by man_ls (guest, #15091) [Link]

Even in the current generation of ultrabooks SSDs and rotating HDDs are split in similar proportions -- HDDs retain the majority in my latest completely unscientific survey. Perhaps in 4 years SSDs will displace HDDs in the high end, and in 8 years in the low end. Then give it some 4 more years to have a reasonable replacement of the existing base, and I will say that a considerable proportion of laptops will have rotating disks until at least 2025. I certainly hope that tux3 reaches the mainline before that.

This may look like a very long time, but please look at the facts. Standard SSDs are 128 GB, while standard HDDs are 500 GB. Apple e.g. overcharges €300 to have a 256 GB SDD, and €800 for 512 GB; street prices for the latter are at least €300~800. I assume the most expensive disks are worth their higher price in reliability, as in HDDs there is no appreciable price difference. People often need those 500 GB and will buy an external HDD if the internal SSD cannot have them, moving the spinning rust outside the main unit but not eliminating it.

Best case, Moore's Law would suggest a 1/2 price drop every 1.5 years, so it would take about 6 years for SSDs to reach current HDD prices. And HDDs are also getting better all the time, so in 6 years we would probably be looking at 1 GB HDDs as the standard. Say in 7.5 years SSDs reach price parity with HDDs; still it would take an upgrade cycle to remove all the legacy HDDs, which we can optimistically set to 4 years. Again, 12 years.

And this is just in the first world...

Displacing HDs

Posted Jan 2, 2013 4:50 UTC (Wed) by daniel (guest, #3181) [Link]

Absolutely agreed about the need to support spinning media well for the next long time. Workstations, personal mass storage, cloud, developing economies, this all adds up to 10 years more of spinning disk optimization. Fortunately, optimizing for spinning media usuaalso optimizes for flash, example: keeping related blocks close together not only reduces seeks, but reduces false sharing on erase blocks, thus reducing write multiplication in the FTL

-- Sent from my Desire Z, Powered by Linux

Displacing HDs

Posted Jan 2, 2013 21:31 UTC (Wed) by cmccabe (guest, #60281) [Link]

I don't think so.

http://arstechnica.com/gadgets/2012/12/ssd-prices-are-low...

Even in 2012, if a laptop didn't have an SSD, it wasn't high-end. Similarly for a database box.

Remember that most desktop users aren't running Linux with a minimal desktop environment. They're running some iteration of Windows. For a lot of these users, adding more cores won't help them go faster, since the giant precompiled binaries they run can't (yet) use all the cores. Adding an SSD always helps, though. That's why users are buying SSDs in droves.

Hard drive prices also never quite recovered from the flooding in Thailand. This shouldn't really be surprising: why should manufacturers plow more investment into a dying technology?

http://www.techspot.com/news/48930-hard-drive-prices-to-r...

Hard drives will remain relevant for a while for people who want to store huge amounts of data. But they'll be a specialty item, like liquid-cooled cases or CRT monitors (remember those?) In 5 to 10 years, talking about hard drives will immediately identify you as an old person, just like reminiscing about those blurry CRT monitors we stared at in our youth.

Displacing HDDs

Posted Jan 2, 2013 22:26 UTC (Wed) by man_ls (guest, #15091) [Link]

Note that I am not discussing the relative merits of SSDs and HDDs, just the hard realities at this point in time. I spent more than €100 on a 20 GB SLC drive because I also think that a good SSD is the best investment for a new computer; but they will not be a true replacement until prices are comparable. Just as when LCD monitors cost over €1000 (or $1000, prices are equivalent), they were not a replacement for CRTs but an expensive luxury.

I see nothing in your links that contradicts what I am saying. The first article acknowledges that SSD prices are about $1/GB, while HDDs are $0.05~0.10/GB -- about 10~20 times lower. The second link states that HDDs will not recover until next year, hardly the time frames we are talking about.

A low end laptop is about €500, and a cheap 500 GB SDD is bound to cost almost that much. Until a comparable SSD costs about 10% of the laptop price it will not be a replacement -- people will just have an external HDD for bulk data. As to the high end, you can redefine it to be "laptop with SSD", but if you look around you will see plenty of €1000+ models with an HDD -- starting with the lower MacBook Pro. There are also many models with a 128 GB SSD, but at 1/4th the capacity they are hardly a replacement.

We are just not there yet, sadly. A breakthrough may come and shorten the time frame, but I don't think that TLCs are it: they will just lower the reliability and may give the whole SSD category a bad name. Not even high-temperature NAND annealing will help as it is (from your first link) at least 10 years off. Perhaps we can agree on a time frame of about a decade until HDDs become a rarity.

Displacing HDDs

Posted Jan 2, 2013 23:02 UTC (Wed) by martinfick (subscriber, #4455) [Link]

But unlike CRTs, spinning disks will likely have a place along side SSDs for a long time since they will likely always provide more storage than SSDs. And when does anyone ever say "I will never need more storage"? Our fish grow to fill our fishbowls with storage more and more. And since spinning disks are getting SSDs built into them, I supect that lone SSDs might actually see a drop before spinning disks do. So while perhaps every laptop will soon have an SSD, those SSD will likely be hybrids so they will also have spinning disks.

Displacing HDDs

Posted Jan 3, 2013 3:32 UTC (Thu) by cmccabe (guest, #60281) [Link]

SSDs don't have to reach price parity with hard drives to displace them. They just have to be the best decision for someone building a new computer. For most users, it's better to redirect some of the dollars you could spend on a giant CPU towards an SSD because that will actually make a difference to performance. In the same way, if you are building a gaming box, you want to buy a reasonable graphics card, rather than a giant CPU and a crummy one. (Personally, I don't play games so much any more, but I know some people who do.)

There was a time when graphics accelerators were unheard-of and everyone just used onboard video. But graphics accelerators didn't have to reach price parity with onboard video to win. They just had to become the best choice for most users. (Nowadays, onboard video is making a comeback due to some excellent chipsets from Intel, but that's another story...)

To sum up: in laptops, phones, music players, and tablets, hard drives are either entirely phased out or nearly there. HDDs will be with us for a while on desktops and servers, but mostly as the new tape drive.

I wonder what kind of performance Tux3 will get on SSD. SSDs do benefit from larger contiugous writes as opposed to small, scattered writes. So to the extent Tux3 can provide that with its deferred I/O mechanism, SSDs will benefit. Of course, SSD firmwares also try to accomplish the same general goal of write coalescing. So there is a risk of burning CPU on something that firmware is already doing.

Displacing HDDs

Posted Jan 3, 2013 7:27 UTC (Thu) by daniel (guest, #3181) [Link]

Just for the record, we're serious about optimizing for SSD as well as HDD. A lot of optimizations apply to both, or at least, one doesn't hurt the other. Eventually there might be tuning parameters, if we hit any optimizations that are mutually incompatible. Optimizing for SSD is made a lot harder because everybody has their own FTL algorithm, we need to guess how they're trying to guess what we're trying to do, if you know what I mean. But writing less total data is always going to be a win on any device, and we think we're doing a pretty good job of that right now.

Displacing HDDs

Posted Jan 3, 2013 22:50 UTC (Thu) by bronson (subscriber, #4806) [Link]

> we need to guess how they're trying to guess what we're trying to do

A perfect description of the emerging FTL insanity. Great quote.

The Tux3 filesystem returns

Posted Jan 2, 2013 7:56 UTC (Wed) by keeperofdakeys (subscriber, #82635) [Link]

Unfortunately, if you power down the drive, then you can no longer read from it. Considering that your RAM is usually smaller then the amount of data you need to access at once, this makes turning the drive off infeasible. The exception are storage drives, which usually don't have many continuous writes anyway.

The Tux3 filesystem returns

Posted Jan 2, 2013 8:43 UTC (Wed) by daniel (guest, #3181) [Link]

Maybe that could be fixed by having a way to preload vast amounts of disk volume to cache. Anyway, I agree, spinning down is useful mainly for volumes that are either completely offline or have long periods of inactivity. Like three of the four disks in my workstation right now, which spend most of their time spun down courtesy of noflushd. BTW, the fourth disk will be changed out for a SSD pretty soon. Not doing so is just cruel and unusual self inflicted punishment.

The Tux3 filesystem returns

Posted Jan 2, 2013 8:47 UTC (Wed) by keeperofdakeys (subscriber, #82635) [Link]

Unless 'cache' is an SSD (which is quite rare), you aren't going to be able to spin down a drive anyway (you also can't control the cache located in most harddrives, since it's for a different purpose). Personally I'm not a fan of this SSD cache phase, waste of an SSD when you could dedicate it to your root OS.

The Tux3 filesystem returns

Posted Jan 2, 2013 21:40 UTC (Wed) by daniel (guest, #3181) [Link]

Note that I'm already getting very effective results out of spinning down my disks. I just take care which data I put on which disk. Spinning down the root volume effectively would be trickier, but doable.

By aggressive caching I meant preemptively reading big linear chunks of volume into host memory, a behaviour we plan to support after more pressing issues are under control. Spindown isn't our immediate priority by any means, but we can probably help make it practical in more situations.

The Tux3 filesystem returns

Posted Jan 2, 2013 11:01 UTC (Wed) by juliank (subscriber, #45896) [Link]

Consider browsers like Chrome that will write to your disk while you are browsing in order to store the current session. This will in normal use only result in writes to the disk, but no reads at all, allowing you to spin down the drive.

The Tux3 filesystem returns

Posted Jan 2, 2013 11:22 UTC (Wed) by keeperofdakeys (subscriber, #82635) [Link]

The session data involves a few things, including a cache. The on-disk cache is read many times, especially when you go to another web-page on a site you are visiting. You are also forgetting all the other things that need to read your disk, like Chrome loading all the dynamic libraries (you may not have everything loaded into RAM at that time), and the rest of your OS.

For safety, I doubt the kernel would allow a device to spin down while there are dirty pages for that disk anyway.


Copyright © 2017, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds