I looked at it with UnixBench (ftp.tux.org:/pub/niemi/unixbench). fsdisk
(4K blocks) and fstime (1K blocks) are much more disk-bound even with 64 MB
of RAM. Could someone take a look with Bonnie and see if there is any
change in Real Disk I/O vs. mid-2.1 series kernels?
I also tried the SMP kernel on UP, which is now extremely close to UP on UP
except in pipe throughput (-45%), pipe-based context switching (-50%),
fsbuffer (256 byte blocks, -30%) and null syscall (-25%). The other 7
scores were all within 10%, and 4 of them were exactly the same. This is a
nice improvement for the SMP overhead, now if pipes can be tweaked a bit
this could get very close.
I also tried it briefly on a Compaq 6000, turning off quite a few nonessential items and could not get it to boot (complained about not being able handle a series of paging requests; this is my work machine and I don't have the exact wording it with me). It fails sometime after detecting the obligatory PS/2 mouse (without which it won't boot under *any* kernel!). 2.1.92pre1 was fine on this machine (Pentium Pro 200, 96 MB RAM, 2 GB Seagate fast narrow AIC 7xxx SCSI, and TLAN Ethernet. I turned off APM altogether and did a complete kernel rebuild, still no dice. My next suspect would be the AIC7xxx driver, if no-one else sees this I'll pin it down next week (as soon as I have time to wait through its glacial POST a few times).
I couldn't reproduce this problem on my much less finicky (and much faster) AMD K6 233, with the same kernel options except Ethernet and SCSI.
--- David C Niemi --- email@example.com --- Reston, Virginia, USA --- "Well the lush separation unfolds you - and the products of wealth push you along on the bow wave of their spiritless undying selves."
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to firstname.lastname@example.org