[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Fragmentation of Journaling FS
> Hi Constantin
>
> Sorry, I'm in doubt with the sharp performace drop at 50% disk usage on a
> xfs filesystem.
Hmmm, you probably need to do some random deleting in between times here,
if I read Constantin's page correctly he is trying to simulate the an
aged filesystem which has had lots of data created and removed over
time - this has the effect of making the free space distribution a lot
more random.
Steve
>
> I made a quick and dirty test running this:
>
> while time cp -a /usr/src/linux/drivers/ /mnt/xxx-`date '+%s'`; do sync; \
> df | grep mnt; done
>
>
> /mnt is a 4GB lvm volume on a 18GB 10000rpm IBM SCSI Disk.
> It's formatted with default mkfs.xfs (no tuning).
> /usr is a LVM volume on this disk too.
> Athlon 650/ 256MB RAM.
> Linux-xfs kernel 2.4.8-pre3 (CVS from 2001-07-31).
> The test was running in multiuser mode with X.
>
> du -ks /usr/src/linux/drivers/
> 73980 /usr/src/linux/drivers
>
>
> Here are the results:
>
> user system elapsed CPU Used Avail. Use%
>
> 0.10 2.98 0:30.25 10% 95196 4094308 3%
> 0.15 2.78 0:29.47 9% 169176 4020328 5%
> 0.14 2.75 0:27.83 10% 243156 3946348 6%
> 0.15 2.86 0:27.04 11% 317136 3872368 8%
> 0.03 3.10 0:26.61 11% 391116 3798388 10%
> 0.07 2.86 0:27.88 10% 465096 3724408 12%
> 0.09 3.04 0:27.26 11% 539076 3650428 13%
> 0.14 3.02 0:27.06 11% 613060 3576444 15%
> 0.10 2.98 0:27.48 11% 687040 3502464 17%
> 0.11 3.14 0:28.07 11% 761020 3428484 19%
> 0.13 3.12 0:28.17 11% 835000 3354504 20%
> 0.12 3.19 0:28.03 11% 908980 3280524 22%
> 0.09 3.27 0:27.71 12% 983024 3206480 24%
> 0.05 3.04 0:27.93 11% 1057452 3132052 26%
> 0.18 3.06 0:28.12 11% 1131816 3057688 28%
> 0.13 3.24 0:28.57 11% 1206244 2983260 29%
> 0.10 3.04 0:28.55 10% 1280608 2908896 31%
> 0.16 3.61 0:28.37 13% 1355036 2834468 33%
> 0.12 3.26 0:28.59 11% 1429400 2760104 35%
> 0.16 3.10 0:29.04 11% 1503844 2685660 36%
> 0.08 3.66 0:29.75 12% 1578192 2611312 38%
> 0.12 3.63 0:29.05 12% 1652604 2536900 40%
> 0.11 3.60 0:29.53 12% 1726968 2462536 42%
> 0.20 3.70 0:29.48 13% 1801396 2388108 43%
> 0.13 3.81 0:29.24 13% 1876096 2313408 45%
> 0.12 3.72 0:29.29 13% 1950908 2238596 47%
> 0.12 3.97 0:29.96 13% 2025720 2163784 49%
> 0.22 3.78 0:29.46 13% 2100532 2088972 51%
> 0.08 3.94 0:30.05 13% 2175104 2014400 52%
> 0.10 3.76 0:30.35 12% 2249084 1940420 54%
> 0.15 3.61 0:30.43 12% 2323240 1866264 56%
> 0.18 3.45 0:29.15 12% 2398116 1791388 58%
> 0.06 4.04 0:29.33 13% 2473056 1716448 60%
> 0.16 3.94 0:31.83 12% 2547996 1641508 61%
> 0.16 3.71 0:34.60 11% 2622920 1566584 63%
> 0.10 4.12 0:30.80 13% 2697876 1491628 65%
> 0.12 4.13 0:29.61 14% 2772768 1416736 67%
> 0.14 3.99 0:30.26 13% 2847708 1341796 68%
> 0.15 3.81 0:29.50 13% 2922632 1266872 70%
> 0.12 3.93 0:29.31 13% 2997572 1191932 72%
> 0.10 4.07 0:29.44 14% 3072512 1116992 74%
> 0.18 4.13 0:33.74 12% 3147468 1042036 76%
> 0.19 4.09 0:36.55 11% 3222424 967080 77%
> 0.16 4.00 0:36.65 11% 3297364 892140 79%
> 0.19 4.50 0:34.12 13% 3372304 817200 81%
> 0.13 4.38 0:37.02 12% 3447244 742260 83%
> 0.06 4.38 0:36.82 12% 3522168 667336 85%
> 0.11 4.21 0:41.77 10% 3597124 592380 86%
> 0.11 4.11 0:38.03 11% 3672016 517488 88%
> 0.12 3.97 0:38.16 10% 3746956 442548 90%
> 0.17 4.33 0:47.37 9% 3821896 367608 92%
> 0.15 4.53 0:47.34 9% 3896820 292684 94%
> 0.16 4.34 0:46.26 9% 3971760 217744 95%
> 0.16 4.30 0:47.54 9% 4046700 142804 97%
> 0.16 4.31 0:49.44 9% 4121640 67864 99%
>
>
> My results looks very resonable for me. A sliding performance degrade with a
> full disk. No performace sharp drop at about 50% usage.
>
> This is my real life experience too.
>
> Is it possible to get your agesystem tool?
>
>
> cheers
>
> utz lehmann
>
>
>
>
> Constantin Loizides [Constantin.Loizides@isg.de] wrote:
> > Hello,
> >
> > I would like to announce the new version of my
> > fragmentation project website at
> >
> > http://www.informatik.uni-frankfurt.de/~loizides/reiserfs/
> [...]
> >
> > Two results of the "agesystem" tool I describe on the page, really are
> > strange and need to be understood. Why is there the sharp performance
> > degrade
> > of XFS and JFS? (the cpu time does not show this behaviour, so it
> > seems to be disk time). Surely more work has to be done, newer versions
> > of the
> > systems to be tested, different setups to be tried. Please note,
> > that agesystem is a misleading term, it doesnot age up to now, it just
> > write to the disk once without deletion of any created file.
> [...]