FileSystem Benchmark

Message

Genkaku · Post by **Genkaku** » Thu Oct 14, 2004 9:20 pm

Yalla.

I'll just post my results of some lil' fs benchmarks:

FS/Test	  *1		*2		*3		*4		*5		*6

ext2	(real)0m2.943s	0m32.818s	2m16.690s	0m16.609s	3m56.574s	(first)15s
	    (user)0m0.002s	0m22.967s	0m0.070s	0m0.009s	3m26.373s	(second)2,5s
	    (sys)0m0.141s	0m1.502s	0m1.649s	0m0.154s	0m15.130s	
	
JFS	  0m3.700s	1m47.886s	5m7.329s	0m50.884s	4m16.316s	13s			
	      0m0.004s	0m23.258s	0m0.104s	0m0.014s	3m28.559s	2-3s		
	      0m0.168s	0m2.147s	0m1.928s	0m0.328s	0m17.222s		

XFS	  0m6.490s	1m2.278s	2m25.033s	0m25.449s	4m3.170s	14s	
	      0m0.003s	0m23.519s	0m0.098s	0m0.030s	3m28.080s	2,5s					
	      0m0.162s	0m3.039s	0m2.872s	0m1.062s	0m16.710s							
												
Reiser4	  0m6.082s	0m33.895s	0m46.481s	0m9.657s	4m8.398s	16				
	      0m0.000s	0m23.529s	0m0.073s	0m0.016s	3m26.641s	2,5s					
	      0m0.160s	0m4.044s	0m3.181s	0m1.843s	0m36.197s							
	  											
ReiserFS  0m0.141s	0m35.667s	2m6.010s	0m0.870s	4m2.503s	14s			
	      0m0.000s	0m23.354s	0m0.095s	0m0.013s	3m27.194s	2,5s						
	      0m0.141s	0m3.190s	0m2.631s	0m0.734s	0m17.756s							
	  											
**1. time cp /usr/portage/distfiles/linux-2.6.8.tar.bz2 /mnt/hda4
**2. time tar xjf linux-2.6.8.tar.bz2
**3. time cp -R linux-2.6.8/ linux
**4. time rm -R linux
**5. cd linux;cp /root/konfa /mnt/hda4/.config;make menuconfig;time make
**6. time of /mnt/hda4/OpenOffice.org1.1.2/program/scalc (first, sekond startup)

Jake · Post by **Jake** » Fri Oct 15, 2004 12:13 am

*1 can't possibly be right. No way is ReiserFS 46x faster than XFS at something XFS should be better than ReiserFS at. Could caching account for such a significant difference?

*2 what FS the tarball was created on matters, did you create a new tarball on each FS to test that FS?

*3 amazing if correct

*4 deleting files is supposed to be reiser4's weakness, so I question your results, although not as much as *1

*5 if you want to see the differences better in this benchmark, use "make allyesconfig && make". Running more jobs would be interesting too (make -jx, where x is >1 and unless you're crazy, <50; normally it's used for SMP systems in which case it's between 1+# of CPUs and 2x# of CPUs)

Genkaku · Post by **Genkaku** » Fri Oct 15, 2004 1:04 am

/usr/portage/distfiles/ was on XFS.

Maybe i'll do some other benchmarks later (yup, creating tarball from kernel source).

I created Reiser4 on /dev/hda4 (above 7GB). And...

-First system startup was awfuly long.
-Second was better, but:
*Firefox, Xorg, kde - such programs were starting 3 times longer than on my XFS partition.
-emerge sync - blazing fast,... caused my system laggy, veru unresponsive for moments.
-emerge whatever is causing same effects as --sync

WTF is wrong with this filesystem ?!

Now I'm going to do Reiserfs on hda4(this hda4 is my test-root), and i'll check what happens.

Jake · Post by **Jake** » Fri Oct 15, 2004 2:09 am

What kernel are you using?

playfool · Post by **playfool** » Fri Oct 15, 2004 9:27 am

I take it this was on the same partition, reformated to "filesystem foo" everytime, to avoid disc placement screws in your timing (data closer to the inner rim is traditionally faster accessed than the outer rim)

waverider202 · Post by **waverider202** » Fri Oct 15, 2004 5:49 pm

all these benchmarks are tainted

here's the thing people need to realize about filesystem benchmarks, linux caches the hard disk. Which means, timing the copying of a file or directory doesn't mean anything cause the read/writes didn't actually happen yet. If you think that isn't true, then try coping a 1.4mb file onto a floppy

It'll take less than a second. Now, try unmounting the floppy, and it'll take some time. Because of Reiser's magic it does with search tree's, it tends to be faster with inode-related opporations. People interpret this as being "faster" than everything else, but in the end, the actual data moving around hasn't happened. Very often, you'll find other filesystems that outperform riser in many instances. Because of the caching, it is very hard to benchmark a fs. The only way for you to really know which one is best for you is to try out each fs for a while, and do what you normally do. Unless you run a server, $5 says you can't tell the difference.

Genkaku · Post by **Genkaku** » Fri Oct 15, 2004 7:44 pm

My kernel is 2.6.9-rc4-nitro1.

Yup, I agree it is best to test some FS by using it for a while. I'm doing so ;]

malloc · Post by **malloc** » Fri Oct 15, 2004 8:09 pm

There's a little point you're missing in your bench and that it makes a hell of a diference (imo at least) wich cpu usage. Reiser and reiser4 are true cpu hogs whilst xfs is not.

Jake · Post by **Jake** » Fri Oct 15, 2004 8:39 pm

malloc wrote:There's a little point you're missing in your bench and that it makes a hell of a diference (imo at least) wich cpu usage. Reiser and reiser4 are true cpu hogs whilst xfs is not.

As flawed as these benchmarks are, kernel compilation is affected (to the point of being bound) by CPU usage, so that issue has been addressed. Sure you wouldn't want to use reiser4 on a slow CPU, but that's the whole point. Usually people have way more CPU power than they need and hard drive access is the bottleneck. Surely we have CPU power to waste with out Athlon64s. And while we're bringing up comparisions other than speed, remember that reiser4 has the advantage of being fully atomic, easy to program with plugins, and stores tiny files amazingly efficiently which comes in handy for the metadata feature other FSs lack.