Ubuntu 9.04 Core 2 Duo (P8400) - SMP slightly unstable

[mwsealey]

After getting through a lot of hard crashes in the beta, the 3.0.0 version works far, far better.

However, at random points I am getting random segfaults of tasks. They are not serious - it is not a system service, but it is happening under high load (for instance, the only time I can reproduce it is to build Angstrom using OpenEmbedded - 2 BitBake threads and 2 make threads from each (this is usually a good balance of a load average of 2.0 to 3.8 on my system). Building successive attempts will fail at completely random stages - rerunning it will succeed the last failed task and the continue to the next, then 3 or 4 tasks down the line will fail. This is not just compiling - it's simple configure script runs, archive extraction (do_unpack, do_distribute_sources) and even in the middle of a real build (linking WebKit).

However I cannot see or find any core dumps or get anything to generate them. It just seems to silently fail, no oops, no panic, no BUG: error in dmesg or other logs, it just kicks out, perhaps with GCC doing an ICE is the most visible change.

Going back down to 1 virtual processor, and everything works beautifully (merely half as fast

[cornelius]

I have the exact same problem with VirtualBox 3.0. With SMP enabled (2 processors), for example, trying to build OpenCog fails at random points, continues when I retry, but fails again at some other point.

I don't know if it's related, but (again with 2 processors) I also get random freezes with VirtualBox using excessive CPU on the host. The freezes usually happen at moments of high load on the guest (like, during make, or when I'm starting a program).

So far, I haven't had these problems when the number of processors is 1.

Logs don't show any errors, as far as I can see. Also, this VM was created with 2.2.4, but I guess that's not a problem.

[mwsealey]

I have the exact same problem with VirtualBox 3.0. With SMP enabled (2 processors), for example, trying to build OpenCog fails at random points, continues when I retry, but fails again at some other point.

I don't know if it's related, but (again with 2 processors) I also get random freezes with VirtualBox using excessive CPU on the host. The freezes usually happen at moments of high load on the guest (like, during make, or when I'm starting a program).

So far, I haven't had these problems when the number of processors is 1.

Logs don't show any errors, as far as I can see. Also, this VM was created with 2.2.4, but I guess that's not a problem.

I installed all my guests again under 3.0.0 to see if it helped, and with or without VBox Guest Additions it's the same way. I also get the freezes but they are nowhere near as annoying (a guest can simply be restarted.. although I did have one major filesystem corruption event this weekend.. I have backups). What annoys me is the completely random segfault behavior - things just crash out, and it's a huge plethora of commands, in many environments (GNOME, runlevel 3, console, JeOS with nothing running and 384MB of RAM, no swap) all when under high CPU load.

To me this sounds like some kind of race condition in the VM... I did not try running a 32-bit guest but I am not sure it would make any difference.

Edit: freezing seems to occur under high load and high DISK ACTIVITY with a single processor, too. The type of controller does not make a difference (I had SATA to start, but have set it back to PIIX4 and then PIIX3 and had the same effect - "rm -rf tmp.3" to clean up my Beagleboard staging directory for example would lock after 2 seconds. VBox.log says nothing, the Guest is unresponsive, but the ACPI Shutdown event is noticed (when I click it, the log says so) and Reset works. I have also tried turning off PAE/NX (not sure what this would do..) and IO-APIC (since it's a single processor) but it is just unstable whatever configuration it's in - it just crashes much, much faster (to the point that enabling 2 processors on my Ubuntu install cannot even mount the root filesystem sometimes - using /dev/sdX and not UUIDs btw - switching back to single processor fixes it as if nothing was ever wrong)

One other thing that is perplexing me; why isn't the full featureset of my CPU being exposed to the guest OS? SSE3 and SSSE3 and a lot of other CPU niceties are completely missing from the point of view of every guest OS. I actually wanted to do some work with these instruction sets in a VM and I can't...

[jbrown]

I am experiencing similar behavior with a Windows 7 host and Windows Server 2003 client running as a build server. All is well with 1 virtual CPU but with multiple virtual CPUs under load programs fail randomly with access violations and the VM ultimately loses the ability the communicate over the network and eventually freezes completely. When running interactively, the VBox window may also become filled with junk and resize itself to a non-sensical size as if the display mode and video memory had become corrupted. At that point it becomes unresponsive and the VM must be powered off and any saved state must be discarded.

Evidently with multiple virtual CPUs under load (like running a build), the VBox client is experiencing memory corruption or I/O corruption until it succumbs.

I wrote about some of this in this ticket which seemed similar at first due to the network problems but has proven to be something else entirely: http://www.virtualbox.de/ticket/4343

Has anyone opened a ticket regarding multiple virtual CPUs being unstable on v3.0.4?