Filesystem for SSD?

Message

Fitzcarraldo · Post by **Fitzcarraldo** » Mon Oct 14, 2019 8:57 pm

mike155 wrote:
a user tested filesystems on four SD cards and F2FS showed increased life up to about 70%.
If I had a SD card, I would also prefer F2FS over ext4.

But we are talking about SSDs, aren't we? SSDs and SD cards are entirely different.

True, but Samsung also targeted SSDs with F2FS:

LWN, October 10, 2012 wrote:Unlike jffs2 and logfs, f2fs is not targeted at raw flash devices, but rather at the specific hardware that is commonly available to consumers — SSDs, eMMC, SD cards, and other flash storage with an FTL (flash translation layer) already built in.

Joo-Young Hwang, Samsung, Feb 22, 2013 wrote:• NAND Flash-based Storage Devices
–SSD for PC and server systems
–eMMCfor mobile systems
–SD card for consumer electronics

•The Rise of SSDs
–Much faster than HDDs
–Low power consumption

https://elinux.org/images/1/12/Elc2013_Hwang.pdf

mike155 · Post by **mike155** » Mon Oct 14, 2019 10:21 pm

Fitzcarraldo, thanks for the document.

The tests shown in the document (starting in page 19) were performed on a eMMC card, not on a SSD.

That's also what the summary (last page) says: F2FS is optimized for mobile flash storages - CAN also work on SSDs.

There's absolutely no doubt that...

F2FS is a good file system for SD cards and USB sticks, almost certainly superior to ext4.
F2FS also works on a SSD

But I haven't seen a single proof or at least evidence yet for the statement that F2FS is better than ext4 on SSDs for the average user.

The special feature of a SSD is the controller - which in fact is a computer of its own. The controller does all the magic that is required to get rid of all the idiosyncrasies and disadvantages that raw flash chips have. And that's why it's so important to buy a SSD with an excellent controller - and not a cheap one that in fact isn't any better than a SD card.

In case you haven't done yet, please read the textbook article that explains how SSDs work internally: http://pages.cs.wisc.edu/~remzi/OSTEP/file-ssd.pdf

Fitzcarraldo · Post by **Fitzcarraldo** » Tue Oct 15, 2019 12:12 am

mike155,

For tests on SSDs (SATA and PCIe using Ubuntu 12.10 Server) comparing F2FS, ext4, Btrfs and NILFS2, see e.g. the following Samsung paper from the Proceedings of the 13th USENIX Conference on File and Storage Technologies (2015): https://www.usenix.org/system/files/con ... er-lee.pdf

You could argue that the tests in that paper do not represent an average user (whatever that is), but their results look interesting enough to consider using F2FS on an SSD instead of ext4 (which I use for my HDDs -- laptops, desktops, HTPC and server -- and really like). I'm sure ext4 works fine on modern SSDs, as you and some others have posted in this thread, but I don't think anyone can tell The_Great_Sephiroth categorically that he should consider ext4 over F2FS. From the various articles and posts I've seen, F2FS looks like a reasonable choice (perhaps a better choice, who knows) for an SSD if he wants to try it.

axl · Post by **axl** » Tue Oct 15, 2019 1:42 am

so little love for xfs

(

axl · Post by **axl** » Tue Oct 15, 2019 1:48 am

well, in as many years as I have as linux user, and a gentoo user, I'll tell you, the least problems I EVER had with anything, any component of any system was XFS.

I use XFS extensively on everything. New raspberry PI4... great... let me recompile the kernel with XFS in it... and other goodies. But the point is, XFS was always a starting point of mine. And for me, xfs didn't kill any drive. In fact it saved some data for me.

It's one of those things you don't think about, but was always there. IN EVERYTHING.

axl · Post by **axl** » Tue Oct 15, 2019 1:58 am

tip. try blktrace and iowatcher I think it's called the new utility to map it out. the older one still in portage doesn't work anymore. not that it's complicated to make a new one. the point is... if you enable CONFIG_BLK_DEV_IO_TRACE=y in kernel and use blktrace you can map out when something reads and writes. and can make really cool graphs.

stuff like this: https://www.youtube.com/watch?v=UQ12PS5x53U

Anyway... that's a cool toy to play with. and yes, that is my video, reading through my ssd. in march 2017.

wouldn't you think if you had the fear that stuff gets over-written too often, and you feared your fs, wouldn't you like to map out what it's doing.

I was curious. So I tested. Funny enough I remained with XFS.

Make up your own test.

PS actually I wasn't reading the disk in the test. I don't remember what I was doing. But there's clear reading and writing.

axl · Post by **axl** » Tue Oct 15, 2019 2:14 am

axl wrote:https://www.youtube.com/watch?v=UQ12PS5x53U

btw, the speed scale which often hits 1000MB/s is from an intel 750 drive. I know that because that was my top dawg in the farm.

but now I have 2 900 series running around. and I didn't graph them yet.

I dont know what people consider a ssd these days... but I love mine:

Code: Select all

[root@sanziana:~]# lspci|grep memory
b3:00.0 Non-Volatile memory controller: Intel Corporation Optane SSD 900P

there she is. in pci near the cpu. I loved the 750 but the 900 is way freaking better.

tholin · Post by **tholin** » Tue Oct 15, 2019 5:00 am

The_Great_Sephiroth wrote:So what about the scheduler? I am under the impression that I should use deadline for an SSD, not CFQ since CFQ was designed to minimize head movement, or so I have read.

CFQ was an legacy IO scheduler and it was removed with the rest of the legacy block layer in kernel 5.0

Goverp · Post by **Goverp** » Tue Oct 15, 2019 9:53 am

AFAIR the current default scheduler is "none", which is supposedly good for SSD's and their ilk (there are no head movements blocking IO).
Otherwise current kernels have a choice of bfq and mq-deadline. IIUC a gross oversimplification is mq-deadline is best for servers (best throughput), bfq for desktops (guarantees some bandwidth for interactive work.

The_Great_Sephiroth · Post by **The_Great_Sephiroth** » Tue Oct 15, 2019 2:54 pm

Thank you, Goverp. I am on the current stable kernel, which is 4.19.72 unless I am mistaken. When Gentoo moves to 5.x, I will be there too! I try to stay in-tree as often as possible. Let Portage deal with my software so I don't have a bunch of manual things to deal with.

Axl, I did consider XFS in the beginning, but for my use-cases it is not up to snuff. On mechanical disks it does not do bit-rot protection or checksums, so while it may be fine, it is not for me, BTRFS and/or ZFS are. I am only now looking into SSDs filesystems and I have settled on F2FS due to the speed improvements it offers as well as extended lifespan of said SSD.

axl · Post by **axl** » Wed Oct 16, 2019 10:29 pm

The_Great_Sephiroth wrote:Axl, I did consider XFS in the beginning, but for my use-cases it is not up to snuff. On mechanical disks it does not do bit-rot protection or checksums, so while it may be fine, it is not for me, BTRFS and/or ZFS are. I am only now looking into SSDs filesystems and I have settled on F2FS due to the speed improvements it offers as well as extended lifespan of said SSD.

xfs does now. mkfs.xfs -m crc=1 is a different fs from -m crc=0. implemented in the last few years. they also added an xfs_fsr which is a defrag utility which is completely lovely.

PS

CONFIG_XFS_ONLINE_SCRUB=y
CONFIG_XFS_ONLINE_REPAIR=y

They show up as a new option under xfs. in kernel. but they do have checksum now.

axl · Post by **axl** » Wed Oct 16, 2019 10:41 pm

The_Great_Sephiroth wrote:I am only now looking into SSDs filesystems and I have settled on F2FS due to the speed improvements it offers as well as extended lifespan of said SSD.

Well, I am speaking from a SSD perspective. I have valuable nvme drives. 3 intel 750 and 2 optane. My personal property. Not the place where I work, not anything like that. MY OWN. And I totally entrusted them to XFS. I am talking about the older 750 drives. I think they are about 4 years old. And still 99% health. And there was no reason to doubt xfs for the new ones, that have like a few weeks of time since they started working.

I totally trust xfs, for both mechanical and ssds. I was never tempted to test f2fs. not even once. why would I?

As far as I know, f2fs (to replace vfat/exfat) was created for mmc cards. By samsung. And some people just ran with it. There's nothing to suggest it's a good idea to keep a rootfs on it. At least I haven't seen anything.

eccerr0r · Post by **eccerr0r** » Thu Oct 17, 2019 2:43 pm

Due to the similar sounding "eMMC" and "MMC" I suspect that the former is simply an on-board version of the latter "embedded". I don't know of wear leveling reliability of eMMC (as far as I know, it's fairly good as witnessed in the longevity of phones, tablets, etc. that use them for primary/root storage) but the removable variants of MMC / SD / USB flash sticks leave something to be desired and f2fs will benefit these.

I'll need to try to set up another more recent Gentoo install on a flash stick (SD/USB) and see how long it lasts before it fails with f2fs... past experience with these it's a bad idea to run Gentoo off of removable flash media on ext2fs, though I do have one Lexar branded USB flash stick that has an old Gentoo install on it. I dared not to emerge @world on it without a PORTAGE_TMPDIR on some other media that I don't need to worry about erase cycles on.

The_Great_Sephiroth · Thu Oct 17, 2019 11:19 pm

I have a 1GB Lexar Firefly drive with FAT32 on it THAT STILL WORKS. How it works is beyond me. I used it new in 2006. This drive is insane. I do sha1sum files on it now and it still seems to work!

Fitzcarraldo · Post by **Fitzcarraldo** » Mon Oct 21, 2019 10:31 pm

Coincidentally, the laptop HDD partition in which I store all my data files is nearly full and I had been wondering what to do for several months. I was not looking forward to replacing the 750GB HDD with a higher-capacity HDD or SSD and having to re-install Gentoo. The 13.3-inch laptop does not have a second drive bay but it does have two mSATA slots for SSDs, so I have just bought a 256GB mSATA SSD module (INDMEM Global Corp., Model: DMMS, 256GB mSATA 3D NAND MLC Flash SSD SATA 3) to provide an extra partition for my data. I installed the SSD today and have formatted it as F2FS to see how the filesystem performs. The SSD will not be as heavily used as the HDD (which uses ext2 on the boot partition, ext4 on the root partition and on the home partition, and NTFS on a partition for data), but it will still be interesting to see how the SSD performs over time using F2FS. I will be mainly copying photos and documents to it, and viewing and editing those files. I will also configure a Cloud client to sync the directories on that drive with my Cloud server.

DigitalCorpus · Post by **DigitalCorpus** » Sat Nov 09, 2019 11:45 pm

The_Great_Sephiroth wrote:So what about the scheduler? I am under the impression that I should use deadline for an SSD, not CFQ since CFQ was designed to minimize head movement, or so I have read.

I generally don't like Phoronix's benchmark suite/results because they're more indicative of enterprise-ish workloads and not enthusiast workloads. That said, they have good stuff sometimes:

Kernel 4.17 Schedulers on [Optane NVMe] SSD
Kernel 4.19 Schedulers on [SATA] SSD
Kernel 4.20 Schedulers on [NVMe] SSD
Kernel 5.0 Schedulers on [120 GB NVMe] SSD

However, I can't find the exported "low_latency" variable in /sys or /proc

Edit: BFQ wasn't selected by default, so nvm...

389292 · Post by **389292** » Sun Nov 10, 2019 10:48 am

I'm running my mining rig from a USB3 flash drive with ext4 for almost 2 years now, and it keeps going. No RO optimizations or anything like that, just regular install.
I was quite surprised actually. My first SSD died after only 1.5 years, it was infamous OCZ Agility. So it depends on many factors, not just your FS.
My current SSD is running ext4 because journaling is actually useful to me because of often black outs.

Website · Post by **Zucca** » Sun Nov 10, 2019 10:57 am

eccerr0r wrote:Due to the similar sounding "eMMC" and "MMC" I suspect that the former is simply an on-board version of the latter "embedded".

That is correct.