On Friday 18 November 2005 00:15, Pierre Patino wrote:
Greetings
I downloaded AIM IX http://sourceforge.net/projects/aimbench (which untars to aim9 directory)
There is actually reaim somewhere on sourceforge too, which is a bit modernized
from sourceforge to compare two motherboards against an older UltraSparcII,600MHz Solaris 8 system. As expected the two newer motherboards, both with SuSE 10.0, Athlon 64 3000+ were much faster except for three tests: Sync Random Writes, Sync Sequential Writes and Sync Disk Copies.
In general AIM 7 and 9 seem to be written for larger disk arrays with many disks. On single disk systems the IO tests tend to be 100% disk bound.
In these three cases, the Sun board was 10 to 30 times faster. The Sun board has an EIDE ATA66 hard drive (40GB). Both Linux boards had 250GB SATA drives with 8MB buffers. The Linux FS is the default reiserfs.
Before I drill any deeper, has anyone got any hints?
These tests are basically disk bound. First keep in mind that while modern disks have much larger bandwidths than older disks, seeks and flush latency are actually not really better than older setups. reiserfs on modern disks does a write barrier on fsync. This means it waits until the disk returns that it has written the data to the platters before returning from the fsync. This is usually not done on older disks and often even impossible. Solaris probably doesn't do that or fails doing it. So it ends up doing much less work, but is also much less crash safe. In most applications it doesn't make that much difference, but these AIM tests are an extreme (and somewhat unrealistic) case. -Andi