on Wed, Sep 17, 2003 at 10:52:15PM +0200, mess-mate ([EMAIL PROTECTED]) wrote: > On Wed, 17 Sep 2003 15:19:00 +0100 > Tom Badran <[EMAIL PROTECTED]> wrote: > > | On Wednesday 17 Sep 2003 14:47, Shyamal Prasad wrote: > | > What kind of application is it that is "way too slow" with ext2? I use > | > ext3 with an 80G drive and it is never slow. ext2 should've been > | > faster. I can recommend ext3 as a good choice, it works great for me. > | > > | > Perhaps you don't have DMA turned on? > | > | If you have an application that needs to pull lots of data really fast (using > | directio) for long periods of time, you really should be using XFS. We tested > | a load of filesystems where i am working and only xfs got the throuput we > | needed (~430MB/s off a raid array) whereas the best ext3 and reiser (might > | not have been reiser4 though) could get was about 350MB/s > | > | Tom > | > Uuuhh, > /dev/hda: > Timing buffer-cache reads: 128 MB in 0.17 seconds =752.94 MB/sec > Timing buffered disk reads: 64 MB in 1.47 seconds = 43.54 MB/sec > mess-mate
RAID is going to speed this considerably.
For a comparison point -- SGI was involved in NSort, a project /
high-end data sorting system which attained a throughput capable of
sorting 1 terrabyte of data in 2.5 hours, in 1997:
http://www.ordinal.com/white/whitepaper.html
It attained this performance through several means:
- 32 processors.
- 8 GiB main memory.
- 559 4 GiB disks, arrayed as 1 system disk, a single 280 disk XLV
volume, and 278 temporary disks.
- A large number of independent controller channels (I remember this
from prior research of the topic, can't find a cite now).
The values attained are modest by today's standards, but the principles
are sound: spread your head movements over as many spindles as
possible, keep your channels clean, and use gobs of memory. You'll get
good performance.
Single-spindle ATA disk access is glacial by comparison.
Peace.
--
Karsten M. Self <[EMAIL PROTECTED]> http://kmself.home.netcom.com/
What Part of "Gestalt" don't you understand?
Defeat EU Software Patents! http://swpat.ffii.org/
pgp00000.pgp
Description: PGP signature
