Benchmarking Filesystems Part ||

Neuro · 2006-01-06 19:17:39

Hi,

Probably most of you here read Slashdot and probably everyone have already read this benchmark article. For those who didn't:

Benchmarking Filesystems Part II, by Justin Piszcz

It looks really suprising. It appears that ReiserV4, which all seem to be so fond of, isn't the best idea for low-CPU machines or laptops (where CPU usage=power). I wouldn't be really suprised if it's benchmarks wouldn't shine on fast machines either.

I'm currently running ReiserV3 (and, as pointed out in the benchmark, it's mounttimes suck), but after reading this article I think of moving FSs. Ext3 seems to have gotten a lot better since last benchmark - too bad dir_indexing wasn't benchmarked. I've never really checked out XFS or JFS, but judging from the article these look really promising. Anyone tried these?

Thoughts/comments/overall discussion follows...

kozaki · 2006-01-06 19:55:25

Thank you for relaying this up to here.

Changes since last benchmarking FS are great to know, indeed

klapmuetz · 2006-01-06 19:56:26

I've been with XFS some years now... I really enjoy it... It's fast, it's good, it's a filesystem alright.

When I get my laptop back, I'm going to check out JFS. It's supposed to be very CPU-friendly... Atleast that's what the benchmark thingie on the LKML pointed out. :-)

Reiser3 never impressed me... It fucked up twice after having a power outtage.
I never had a single fuckup with XFS... And I was really mean to it over the years. :-D

Like always: YMMV.

Neuro · 2006-01-06 20:09:51

klapmuetz wrote:

When I get my laptop back, I'm going to check out JFS. It's supposed to be very CPU-friendly... Atleast that's what the benchmark thingie on the LKML pointed out. :-)

Do you still have a link to that message? I read LKML from time to time but apparently I've missed that one.

klapmuetz · 2006-01-06 20:29:23

Neuro wrote:

klapmuetz wrote:
When I get my laptop back, I'm going to check out JFS. It's supposed to be very CPU-friendly... Atleast that's what the benchmark thingie on the LKML pointed out. :-)
Do you still have a link to that message? I read LKML from time to time but apparently I've missed that one.

Not the newest benchmark, but I found it...
http://kerneltrap.org/node/715

While looking for it, I found another benchmark...
http://fsbench.netnation.com/

elasticdog · 2006-01-06 21:06:51

As a reader who saw this posted on both Slashdot and Digg, many commenters had good points on how these numbers don't mean much. The benchmarks are being performed on old hardware (500MHz I believe) which can definitely hurt Reiser's numbers since it's designed to be more CPU intensive, the tests don't account file fragmentation over time, it's hard to say how these would scale, etc.

There are many holes here, which aren't obvious if you're just doing a quick glance at the data, but the numbers are interesting to look at none the less.

Gullible Jones · 2006-01-06 21:13:39

Here are my findings. Again, between hardware and users' needs, YMMV. My results for XFS, for example, probably won't apply to you if you have a fast SATA or SCSI drive, and you'll find ext3 grossly lacking if you're dealing with filesystems of petabyte size (e.g. systems that handle data from particle colliders).

- JFS is fast, CPU efficient, and generally rocks. It's not quite as good with lots of small files as ReiserFS, but very close, and appears to have faster read speeds.

- Indexed ext3 is very very very fast for just about everything... except pacman. It seems, for some reason, that the pacman "database" gets fragmented a bit faster on ext3 than on other filesystems. The slight sluggishness of pacman isn't a showstopper though, and pacman-optimize usually takes care of it for a while. Other than that, though, indexed ext3 is very good.

- Non-indexed ext3 is very slow. My general recommendation is to never use ext3 without directory indexing.

- ReiserFS is pretty good. There's supposedly a performance hit from tail packing, but I've never seen increased performance with tail packing turned off. I do have a major complaint about it though: journal replay on mount, which takes forever for large volumes.

- XFS is incredibly slow, and causes a lot of hard drive grinding, especially when dealing with lots of small files - although for me, it doesn't perform well with large ones either. I've heard tell of ways to make it perform better with small files, but never found details on them, and I'm betting they're quite mythical; at any rate, I don't think it would make a difference, as this filesystem seems to disagree with my HDDs.

- Ext2 (nonjournalled) should NEVER be used on hard drives. I learned this in a rather irritating way, when my ext2 boot partition became corrupted when I accidentally hit the power switch while my machine was still on, and a kernel panic message greeted me on the next boot. There's a good reason to use journalled filesystems, folks!

klapmuetz · 2006-01-06 21:34:01

Gullible Jones wrote:

- Ext2 (nonjournalled) should NEVER be used on hard drives. I learned this in a rather irritating way, when my ext2 boot partition became corrupted when I accidentally hit the power switch while my machine was still on, and a kernel panic message greeted me on the next boot. There's a good reason to use journalled filesystems, folks!

I always had a ext2 /boot partition...
Obviously it is never mounted, except for if I'm going to install a new kernel.]

Working flawlessly... Other than that, interesting post. Filesystem is probably a topic that is really YMMV'ish. :-)

T-Dawg · 2006-01-06 21:34:47

- Non-indexed ext3 is very slow. My general recommendation is to never use ext3 without directory indexing.

I wish I could try this. I recently reinstalled arch switching to ext3 from reiser and noticed a difference in speed in favor of ext3. Can anyone tell me approximently how much faster dir_index is compared to the default?

Gullible Jones · 2006-01-06 23:29:57

I've seen people say that indexed ext3 is comparable to Reiser4, but never having used Reiser4, I wouldn't know. It seems to me that read speeds are a bit faster than with JFS.

Gullible Jones · 2006-01-07 05:07:45

With regard to the OP: those benchmarks are very interesting... Not in the least because Reiser4 seems to show markedly inferior performance for a lot of things. From the looks of it, JFS is still the best of those tested: performs well under most tests, and is much more CPU-efficient than other filesystems. It looks as though the only really big improvement in Reiser4 over Reiser3 is read speeds.

(The mount time thing is interesting, BTW. I knew that both versions of ReiserFS take a while to mount, but I didn't realize that the scaling was that bad... Hours to mount a large RAID volume?! That is ridiculous.)

Edit: hmm... I notice that they used unindexed ext3. That's definitely going to skew the results a bit.

demonus · 2006-01-07 09:18:06

as for ext3 with dir_index option, I'm having some problems recently, especially when closing system with shutdown -h say from xterm in kde, mostly it results in broken fs, and fsck reporting wrong node sizes etc. none of these problems were present when using ext3 without dir_index

Neuro · 2006-01-07 09:28:42

I'm using ext3 dir_indexing on my music partition. Was using FAT32 before and the listing of all directories took ages. Well, maybe not ages but was really noticeable . Then I switched to ext3 (using it under Windows too) and things improved dramatically, yet I was still able to feel the directory listing. With dir_indexing it's blazingly fast.

AFAIK dir_indexing only improves directory listing/searching but also reduces directory writes a bit (since the file must be also written to the directory index). It's really helpful in directories with loads of files.

kozaki · 2006-01-07 12:13:33

As for Ext2 (nonjournalled) that "should NEVER be used on hard drives", I'll go a different way.
I use it on 3 different partitions :
- /tmp, 5 GB mainly for writing ISOs as it takes barelly a few seconds to write NNN MB (a couple of seconds to write a 150 MB ISO file) . It's cleaned at every boot. So why use journalized FS ?
- biggest files (mostly Audio/Video) on an USB external HDD (backup external HDD's partition comes with JFS)
- separate /boot partition, 100 MB, that I can bypass with a Grub-on-a-floppy

Other than that I go with ext3 for main partition & ReiserFS /www partition to handle thousands of small files (I consider shifting to ext3 and benchmark it after reading that FS bench test).

main HDD is a SATAII Hitachi 80 GB
external USB HDD is an IDE Maxtor 160 GB with up to 70 GB partitions

Now I know why my small Toshiba Portege 7200 Series, 650MHz, will stay with ext3

pikass · 2006-01-07 14:16:54

klapmuetz wrote:

I've been with XFS some years now... I really enjoy it... It's fast, it's good, it's a filesystem alright.

Second that. XFS is the best tradeoff for most tasks. (imho)

Gullible Jones · 2006-01-07 14:58:45

IF you have a fast hard drive. I'm guessing that 5000 RPM hard drives with 2 MB caches don't really go with that filesystem, judging from my experience.

tomfitzyuk · 2006-01-07 15:08:22

I'm currently using ReiserV3 and the mount times are terrible, booting up is tremendously slowed down due to this. I'm thinking of moving to JFS or XFS.

smoon · 2006-01-07 15:49:03

tomfitzyuk wrote:

I'm currently using ReiserV3 and the mount times are terrible, booting up is tremendously slowed down due to this. I'm thinking of moving to JFS or XFS.

I just moved all my Reiser-partitions to ext3 last night since the long mount-times were really annoying. So far I'm happy I finally did the switch. But maybe I should test JFS or XFS sometime as well.

kensai · 2006-01-07 15:55:07

I've enjoyed JFS a lot, and if you see it wins cpu utilization in almost every if not on every test. So it is CPU friendly, no wonder why my cpu always asked me to make a / JFS partition.

whargoul · 2006-01-07 21:08:10

So, which is best? XFS or JFS? I'm thinking of swithing from ReiserV3, since it is slow as hell.

Gullible Jones · 2006-01-07 21:33:47

XFS is fine if you have a SATA drive or deal with big files all the time. Just be warned, it will be very sluggish if you don't have a fast hard drive.

JFS is a good all-purpose filesystem, and excels in CPU efficiency. It's about as fast as ReiserFS for most things, and doesn't require ridiculous mount times.

kensai · 2006-01-07 21:46:27

whargoul wrote:

So, which is best? XFS or JFS? I'm thinking of swithing from ReiserV3, since it is slow as hell.

If you love your CPU and HardDrive disk JFS is the better one.

Cotton · 2006-01-07 22:47:48

The advantage for me of using reiser3 over ext3 under lvm is that it can be resized on the fly, ie without umounting the partition.

whargoul · 2006-01-08 01:10:45

Gullible Jones wrote:

XFS is fine if you have a SATA drive or deal with big files all the time. Just be warned, it will be very sluggish if you don't have a fast hard drive.
JFS is a good all-purpose filesystem, and excels in CPU efficiency. It's about as fast as ReiserFS for most things, and doesn't require ridiculous mount times.

JFS sounds fine since I'm (still?) using an ATA hdd. Thanks.

Gullible Jones · 2006-01-08 01:34:03

Welcome. And don't worry, you're not hopelessly outdated, I think most of us are still on IDE.

Arch Linux

#1 2006-01-06 19:17:39

Benchmarking Filesystems Part ||

#2 2006-01-06 19:55:25

Re: Benchmarking Filesystems Part ||

#3 2006-01-06 19:56:26

Re: Benchmarking Filesystems Part ||

#4 2006-01-06 20:09:51

Re: Benchmarking Filesystems Part ||

#5 2006-01-06 20:29:23

Re: Benchmarking Filesystems Part ||

#6 2006-01-06 21:06:51

Re: Benchmarking Filesystems Part ||

#7 2006-01-06 21:13:39

Re: Benchmarking Filesystems Part ||

#8 2006-01-06 21:34:01

Re: Benchmarking Filesystems Part ||

#9 2006-01-06 21:34:47

Re: Benchmarking Filesystems Part ||

#10 2006-01-06 23:29:57

Re: Benchmarking Filesystems Part ||

#11 2006-01-07 05:07:45

Re: Benchmarking Filesystems Part ||

#12 2006-01-07 09:18:06

Re: Benchmarking Filesystems Part ||

#13 2006-01-07 09:28:42

Re: Benchmarking Filesystems Part ||

#14 2006-01-07 12:13:33

Re: Benchmarking Filesystems Part ||

#15 2006-01-07 14:16:54

Re: Benchmarking Filesystems Part ||

#16 2006-01-07 14:58:45

Re: Benchmarking Filesystems Part ||

#17 2006-01-07 15:08:22

Re: Benchmarking Filesystems Part ||

#18 2006-01-07 15:49:03

Re: Benchmarking Filesystems Part ||

#19 2006-01-07 15:55:07

Re: Benchmarking Filesystems Part ||

#20 2006-01-07 21:08:10

Re: Benchmarking Filesystems Part ||

#21 2006-01-07 21:33:47

Re: Benchmarking Filesystems Part ||

#22 2006-01-07 21:46:27

Re: Benchmarking Filesystems Part ||

#23 2006-01-07 22:47:48

Re: Benchmarking Filesystems Part ||

#24 2006-01-08 01:10:45

Re: Benchmarking Filesystems Part ||

#25 2006-01-08 01:34:03

Re: Benchmarking Filesystems Part ||

Board footer