Semi-solved: Fedora 13 I/O Freeze

[joekrahn]

SOLVED??
Based on a copying some flags known to work around a CentOS kernel problem, I tried kernel flags:
nmi_watchdog=0 elevator=deadline nolapic_timer clocksource=acpi_pm

That system is now stable. I tried nmi_watchdog=0, and it froze. I then tried just "nolapic_timer", and that alone seems to fix it, but I have not done sufficient testing to be sure. It also makes sense that timers can be a big source of problems in virtual hardware, and that different kernels will be more or less sensitive to timer errors. So, I suspect that keeping the "clocksource=acpi_pm" option may also be helpful. I have not checked for clock drift.

Good luck,
Joe Krahn

[fixedwheel]

I then tried just "nolapic_timer", and that alone seems to fix it, but I have not done sufficient testing to be sure.

with just "nolapic_timer clocksource=acpi_pm" my 64bit fedora 13 crashed at boot, then i found

# nolapic_timer can be useful on i386; on x86_64 this option is called noapictimer

giving just "noapictimer" did it: byte-unixbench ran ok on 4 CPU, much less clock drift than before. -edit- hmm... now guest time is a bit too slow on idle, the other way it was much too fast with CPU load

[joekrahn]

I have been running 6 core Fedora 13 guest with option "nolapic_timer", with many hours under a heavy load, and it is very stable, but I have not checked clock drift. There are several threads discussing work-arounds for APIC timer problems, which should give some suggestions for kernel options to try.

[fixedwheel]

now my added parms for fedora 64bit are: selinux=0 noapictimer nohz=off highres=off

byte-unixbench ok
no clock drift

cheers!

[joekrahn]

Kernel docs say that `nolapic_timer' is for i386, and `noapictimer' is for x86_64. From kernel 2.6.34 source code, they are now exactly the same.

I am now using: noapictimer nohz=off highres=off
There are still some occasional errors, like no keyboard response typing the password for X11, and once hung when the VBox service started. Once running, it seems very stable.

fixedwheel: Is the selinux=0 relevant, or did you just include it for completeness?

Joe

[fixedwheel]

Is the selinux=0 relevant, or did you just include it for completeness?

not relevant just included for completeness