Ext4 hacker Ted Ts'o converts his laptop
So far Ive found one bug as a result of my using ext4 in production (if delayed allocation is enabled, i_blocks doesnt get updated until the block allocation takes place, so files can appear to have 0k blocksize right after they are created, which is confusing/unfortunate), but nothing super serious yet. I will be doing backups a bit more frequently until Im absolutely sure things are rock solid, though!"
Posted Jun 30, 2008 23:51 UTC (Mon)
by kirkengaard (guest, #15022)
[Link]
Posted Jul 1, 2008 5:19 UTC (Tue)
by avik (guest, #704)
[Link] (33 responses)
Posted Jul 1, 2008 5:55 UTC (Tue)
by proski (subscriber, #104)
[Link] (1 responses)
Posted Jul 1, 2008 6:04 UTC (Tue)
by avik (guest, #704)
[Link]
Posted Jul 1, 2008 6:13 UTC (Tue)
by Zorggy (guest, #51397)
[Link] (29 responses)
Posted Jul 1, 2008 7:23 UTC (Tue)
by Camarade_Tux (guest, #51944)
[Link] (12 responses)
Posted Jul 1, 2008 9:11 UTC (Tue)
by Zorggy (guest, #51397)
[Link] (2 responses)
Posted Jul 2, 2008 18:19 UTC (Wed)
by oak (guest, #2786)
[Link] (1 responses)
Posted Jul 2, 2008 19:00 UTC (Wed)
by martinfick (subscriber, #4455)
[Link]
Swap is one use-case for disk with which SSD doesn't deal that well, I
guess computers of future still need swap...
Uh, probably not. Why bother swapping to an expensive medium when swap does not need to be persistent, just add more RAM. You could even devise slower RAM based non persistent drives if you were worried about cost. Swap should die if spinning disks die. The whole point of swap is that in theory it is cheaper than RAM, not that is is persistent. Makes me wonder why no one is building a cheap non persistent "swap" drive yet, do they?
Posted Jul 1, 2008 9:15 UTC (Tue)
by rvfh (guest, #31018)
[Link]
Posted Jul 1, 2008 10:59 UTC (Tue)
by endecotp (guest, #36428)
[Link]
Posted Jul 1, 2008 23:03 UTC (Tue)
by k8to (guest, #15413)
[Link] (5 responses)
Posted Jul 2, 2008 2:52 UTC (Wed)
by jwb (guest, #15467)
[Link] (1 responses)
Posted Jul 3, 2008 22:36 UTC (Thu)
by liljencrantz (guest, #28458)
[Link]
Posted Jul 2, 2008 9:21 UTC (Wed)
by drag (guest, #31333)
[Link] (2 responses)
Posted Jul 2, 2008 19:27 UTC (Wed)
by nix (subscriber, #2304)
[Link] (1 responses)
Posted Jul 3, 2008 22:39 UTC (Thu)
by jwb (guest, #15467)
[Link]
Posted Jul 2, 2008 6:37 UTC (Wed)
by man_ls (guest, #15091)
[Link]
OTOH 3.5" disks, which have different constraints as to power and energy savings, will draw quite more power (a quick Google search shows peak consumption of 6-13W, instead of 2-4W for 2.5" disks). Meanwhile an SSD disk has the same basic consumption, be it 3.5" or 1.8", as the article shows. So there is a lot of potential savings even in the high end.
Posted Jul 1, 2008 7:42 UTC (Tue)
by trochej (guest, #35052)
[Link] (2 responses)
Posted Jul 1, 2008 7:57 UTC (Tue)
by zzxtty (guest, #45175)
[Link]
Posted Jul 1, 2008 8:51 UTC (Tue)
by Zorggy (guest, #51397)
[Link]
Posted Jul 1, 2008 12:55 UTC (Tue)
by SEJeff (guest, #51588)
[Link]
Posted Jul 1, 2008 16:31 UTC (Tue)
by zooko (guest, #2589)
[Link] (10 responses)
Posted Jul 1, 2008 17:52 UTC (Tue)
by rriggs (guest, #11598)
[Link] (9 responses)
Take a look at the site you just posted on. What do you think the "L" stands for. Given the context, ZFS is not ready. It will never be bootable.
Nexenta isn't ready either, but that's just my opinion.
Posted Jul 1, 2008 18:26 UTC (Tue)
by zooko (guest, #2589)
[Link] (1 responses)
Posted Jul 1, 2008 20:14 UTC (Tue)
by martinfick (subscriber, #4455)
[Link]
Posted Jul 1, 2008 20:15 UTC (Tue)
by proski (subscriber, #104)
[Link] (6 responses)
Posted Jul 1, 2008 21:32 UTC (Tue)
by rriggs (guest, #11598)
[Link] (5 responses)
Posted Jul 3, 2008 21:44 UTC (Thu)
by Los__D (guest, #15263)
[Link] (4 responses)
Posted Jul 4, 2008 19:14 UTC (Fri)
by k8to (guest, #15413)
[Link] (3 responses)
Posted Jul 4, 2008 20:17 UTC (Fri)
by nix (subscriber, #2304)
[Link]
Posted Jul 5, 2008 8:09 UTC (Sat)
by Los__D (guest, #15263)
[Link] (1 responses)
Posted Jul 5, 2008 8:09 UTC (Sat)
by Los__D (guest, #15263)
[Link]
Posted Jul 3, 2008 8:05 UTC (Thu)
by jengelh (guest, #33263)
[Link]
You seem running after the future, Avik, never satisfied with what is existing right now. But look, EXT4 is in the mainstream and will be soon usable even if it cannot directly be compared with BTRFS or ZFS. Extents, delayed allocation, atomic quota switch (i.e. no time window and hence races between mounting the fs and running quotaon), bulk quota scanning, and, on the implementation side, B+ tree, 16 Exabytes volume size, 8 Exabytes file size -- someone remind me of how long XFS has been around, thank you. Ext4 is like how many years late?
Posted Jul 2, 2008 8:40 UTC (Wed)
by oblio (guest, #33465)
[Link]
Ext4 hacker Ted Ts'o converts his laptop
Hooray for the project having reached "eating your own dogfood" stage!
Ext4 hacker Ted Ts'o converts his laptop
I would have liked it much more if he converted his laptop to btrfs. I hate to be dismissive
of the ext4 developers' work, but the fact is that ext4 is already obsolete (and was obsolete
even before development started). Infant btrfs already supports features that ext4 will never
have, and provides the groundwork for even more interesting features and the potential for
great performance.
Considering that Ted Ts'o is one of the ext4 developers, I'd rather see him testing his own code rather than someone else's.
Ext4 hacker Ted Ts'o converts his laptop
Ext4 hacker Ted Ts'o converts his laptop
I knew that; what I meant was that I'd like to see Ted convert his laptop as well as his
development efforts to btrfs.
Again, it isn't very polite to dismiss the huge amount of work that has gone into ext[234],
but I do feel the design was outdated ten years ago, let alone now.
Ext4 hacker Ted Ts'o converts his laptop
You seem running after the future, Avik, never satisfied with what is existing right now. But
look, EXT4 is in the mainstream and will be soon usable even if it cannot directly be compared
with BTRFS or ZFS. However those last fs are far from being ready for production! And if you
really look at the future, standard disks are going to disappear for SSD, which need other
kind of fs.
I'm very happy to hear from EXT4, better than EXT3 for bigger partitions, maybe in production
next year?
Ext4 hacker Ted Ts'o converts his laptop
I still don't get how people imagine everything is about to use SSDs. SSDs die after far less
write cycles than hard drives.
Quoting wikipedia :
"Limited write cycles flash-memory storage will often wear out after 300,000-500,000 write
cycles[citation needed]"
(but citations are easy to find)
This page seems to be a good reference : http://wiki.eeeuser.com/ssd_write_limit
Basically, it states eee user won't have problems. But well, eee users aren't server users or
even regular users.
Compile anything and you're damaging your hard drive (kernel, webkit). Download something
(let's say a Debian iso) and you're facing the same problem. Defragment your drives and you're
dead (vista does this on a regular basis by default). In fact, even the eee users may have
this problem if their browser flushes things to disk too frequently.
So, no, I don't think those are going to be mainstream soon or so I hope.
(and SSDs draw more power than regular hard drives atm :
http://www.tomshardware.com/reviews/ssd-hdd-battery,1955....)
Sorry for the off-topic but I couldn't stand any longer. ;)
Ext4 hacker Ted Ts'o converts his laptop
Of course I did not explained that SSD would replace disks next year, but it is a fast
evolving technology. The example you show is a bit biased as the EeePC is a very low-cost
computer, not integrating the cutting-edge technology, far more expensive.
But you are right, there are some limiting constraints that must be taken into account when
using a file-system. There are some specialized ones such as JFFS2, YAFFS, LogFS or UBIFS
which are far better when managing SSD compared to EXTx, specialized in rotating disks.
And about the consumption, I can read from the article you provided: "we received confirmation
from two vendors that many flash devices dont feature power saving mechanisms yet". That
could be a path for huge improvements in a near future.
Ext4 hacker Ted Ts'o converts his laptop
Regarding JFFS2, it scales *very* badly. I think OLPC has done some work
for it to be usable with 1/2GB partition sizes, but basically JFFS2 is
obsolete technology with today's disk sizes.
I've heard that the larger SSD disks might support less write-cycles
(there's more space over which to spread the writes so it could be less of
an issue for normal desktop users), any idea whether this is true?
Swap is one use-case for disk with which SSD doesn't deal that well, I
guess computers of future still need swap...
Ext4 hacker Ted Ts'o converts his laptop
Ext4 hacker Ted Ts'o converts his laptop
If you read papers from people like Sandisk, they basically say that current SSDs will outlast
your computer. It is not [virtually] illimited in write cycles as HD are, but it is enough for
a normal _desktop_ user.
And it will only get better...
And HDs die in other ways (I lost 5 so far at home, out of about 15)...
Ext4 hacker Ted Ts'o converts his laptop
Like it says, CITATION NEEDED.
Magnetic disks fail too. They've got moving parts in them!!!!! Maybe you've never had a
magnetic disk fail, but I bet you know someone who has.
I've been using exclusively flash storage for a couple of years now and it has been very
reliable. Yes, I compile things, download things and so on; in fact I spend most of my time
doing exactly that, on flash drives. I have had one USB "thumb drive" fail, but it was a
cheap one and I don't believe it was anything related to the write limit; none of my IDE
drives or modules has ever had a problem.
As for the power consumption, I have found my drives to take at least an order of magnitude
less than magnetic disks.
The greatest benefit for me, however, is that they are silent.
Ext4 hacker Ted Ts'o converts his laptop
Ugh, this useless old red herring.
The article shows that if your disk is under continuous writes of megabytes a second without
pause that it will still last for YEARS.
Some server deployments may be of this nature. NO desktop deployments are. If you like I
could monitor iostat data for my workstation that acts a s a fileserver, does bittorrent
periodically, on which I compile software, and develop software and so on. I can assure you
that the disks are idle over 80% of the time.
Ext4 hacker Ted Ts'o converts his laptop
Obviously you've never used tracker ;)
Ext4 hacker Ted Ts'o converts his laptop
Under regular load, 90 % of the IO done by tracker is reading. When tracker hangs and keeps
rereading the same file over and over, 100 % of the io is reading. SSDs can handle it just
fine.
Ext4 hacker Ted Ts'o converts his laptop
Well traditional harddrives have write limits also. And, like any computer person, have had
more then one harddrive fail prematurely.
And, more importantly for the direction that PCs are going.. when you drop your laptop on it's
corner you won't crash the read heads on a solid state disk.
AND, SSDs for most purposes are much much faster. Low seek times, you see.
Ext4 hacker Ted Ts'o converts his laptop
But the write times are lousy. Really, really lousy.
Ext4 hacker Ted Ts'o converts his laptop
The flash SSD market is very segmented. A lot of the disk-form-factor stuff out there has
mediocre performance. However there is a company with a product on the market that's about
100-200x faster than disk on both reads and writes, both random and sequential, and especially
on mixed read/write loads. The pricing is about $25/GB which isn't all that bad when you
break it down.
The article on Tom's Hardware says that power consumption of a high-end SSD disk can be worse than for a 2.5" high-end laptop disk; but the comparison cannot be generalized. It only compares disks with very high performance and steep prices. Cheaper, slower SSD drives may easily draw less power, and they even benchmark a Sandisk unit that shows it.
Speaking about laptops
Ext4 hacker Ted Ts'o converts his laptop
I don't know what you mean by "zfs is not ready for production", since I've seen quite a large
number of big production environments migrate to ZFS with a success. And we're talking about
constant data munging, not some stale one-time-per-month backup activity.
Ext4 hacker Ted Ts'o converts his laptop
Using zfs here, a good 30-40 TB of space, in regular use, so far so good. We are only using
concat, not raid Z, we have proper raid hardware underneath. The import/export ability is very
nice, I really couldn't believe it worked when i tried it (hot swapped two disks between two
machines), it was a pleasure to rip up my contingency plans! Quota and reservation handling is
great, it really does simplify things for our users, it's just a shame it cant be used as a
cluster filesystem! (me, demanding? never!).
Ext4 hacker Ted Ts'o converts his laptop
Sorry, I did not mentioned "ZFS ready for Linux" that is, using Fuse, because EXT[234] is used
on Linux. If ZFS is really used in production on Linux, I am not aware of...
Ext4 hacker Ted Ts'o converts his laptop
btrfs has code that makes it run better on ssd fyi. It does not need "A new filesystem".
Ext4 hacker Ted Ts'o converts his laptop
What do you mean ZFS isn't ready? ZFS is great. Just get http://nexenta.org and install it.
It will install the root partition ZFS. If you have two disks you can do RAID-Z mirroring
across them. Also, nexenta has the beautiful idea of "apt-clone". "apt-clone" is a wrapper
around apt-get which does the normal apt-get install but also uses ZFS snapshots to give you
the ability to roll back to the previous state.
Ext4 hacker Ted Ts'o converts his laptop
What do you mean ZFS isn't ready?
LWN isn't just about Linux
Oh, I see. When the earlier poster said that ZFS wasn't ready he meant that in the context
that he runs only Linux.
However, in general I don't think that LWN.net excludes non-Linux topics.
Many, if not most, of the readers of LWN.net appear to be people who, like me, care a lot
about freedom and openness and good engineering, but who do not feel antagonism or exclusivity
towards other free and open operating systems than Linux.
Indeed, I've noticed that while LWN.net's coverage of Linux is excellent -- very detailed and
up-to-date and well informed and well written -- much of the content on LWN.net is not
Linux-specific. (While the Kernel page is generally all-Linux all the time, the Development
page rarely has anything Linux-specific on it. Both of these are as they should be.) I've
also observed that LWN editors choose to include other free and open operating systems in
their articles.
That's why, when I pronounce "LWN" in my mind, it either comes out as "Libre Weekly News", or
more commonly as simply "Ell Doubleyou Enn".
Regards,
Zooko
LWN isn't just about Linux
I don't think that exclusion was the issue. The issue is that it would be silly for a linux
developer to simply drop developing robust filesystems for Linux simply because Solaris
happens to have one. :)
Being bootable is not equivalent to being ready. In fact, the term "bootable" is not well defined. You can have a small ext2 partition for the kernels and the bootloader. You can have an initial ramdisk with fuse and ZFS support.
Ext4 hacker Ted Ts'o converts his laptop
Ext4 hacker Ted Ts'o converts his laptop
You just let me know when you have that running in a production environment where downtime
costs are not trivial.
Ext4 hacker Ted Ts'o converts his laptop
I thought most production environments made sure to have /boot on a separate (ro mounted)
partition...
Ext4 hacker Ted Ts'o converts his laptop
The reason for this is historical.
On really old Unixes, filesystems *would* become mangled with some regularity. You brought
systems down and fscked them regularly in order to discover where, not if, things had gone
wrong, and restore them. A small /boot type partion was relevant in order to decrease the
likelyhood that a serious error would occur on the boot filesystem.
On Linux, which has had reliable filesystems since at least ext2fs became widely available
(1994 or so?) this wasn't really an issue. However, because Linux boots with assistance from
a poorly maintained ROM interface, which tends to often fail to provide rom-call access to the
entire disk, a /boot filesystem can make sense for the general case, since it doesn't really
cause harm when it is not needed.
However with the advent of generally available 48 bit (or whatever, the larger addressing
version) of LBA in BIOS implementations, we haven't had to deal with this problem for some
years. Current linuxes seem to not require a /boot partition, that I have dealt with. I
certainly haven't configured any production systems with a /boot partition for a long time.
Ext4 hacker Ted Ts'o converts his laptop
There's another reason to use /boot: if your boot program won't boot off
the filesystem or block device that your main OS is on.
A major example of this is any of the more complicated RAID setups. Most
of my systems are LVM atop RAID-5 or RAID-6, sometimes NBDed to other
remote systems (using write-mostly to avoid incredible sloth). Obviously
you want / to be RAIDed, because if /lib or /etc dies you're in big
trouble: but it's highly unlikely that LILO or whatever could boot off
that. So you make a smaller RAID-0 /boot, with an initramfs or whatever so
that it can bring up RAID and LVM, and put your kernels on there. Bingo.
Ext4 hacker Ted Ts'o converts his laptop
I do it simply to make sure that I don't do nasty things with my kernel by accident.
Anyway, I still don't see the any problems in relation to the OP's mentioning of downtime
problems.
Ext4 hacker Ted Ts'o converts his laptop
Ehm, that came out a bit funny, but I guess you understand.
Ext4 hacker Ted Ts'o converts his laptop
Ext4 hacker Ted Ts'o converts his laptop
From: kt4@prism.gatech.EDU (Ken Thompson)
"viewpoint may be largely unrelated to its usefulness. Many if not
most of the software we use is probably obsolete according to the
latest design criteria. Most users could probably care less if the
internals of the operating system they use is obsolete. They are
rightly more interested in its performance and capabilities at the
user level."
Source: http://oreilly.com/catalog/opensources/book/appa.html,
You could take this to heart.
What is "obsolete" (who decided?, according to which criteria?), and why should we care about
it? Remember, the wheel is a several thousand year old design, and that doesn't mean it's
obsolete :)