tart
tart copied to clipboard
After updating to 0.19, anyone seeing GPU tearing that ultimately crashes the VM when scrolling in Safari?
Did you also upgrade something else? Like Xcode or macOS? Or tried a different VM?
I didn't change anything else and haven't tried any other VMs yet.
Interesting. Which version of macOS your VM has? Also which version of macOS you host has?
I saw a GPU crash only once a while back on macOS Monterey. VM UI was unresponsive but I was able to SSH in and top
was showing DumpGPURestart
just consuming everything. I created a Feedback Assistant FB10025528 and was able to chat about it with Apple folks during WWDC. Their suggestion was to use Ventura since they did a lot for the paravirtualized GPU support including drivers in Ventura itself.
But it still look interesting that you started experiencing it after an upgrade. 0.19.0
version doesn't have any changes related to VM configuration except that 0.19.0
was built with a new Xcode Beta 5. What kind of workflow you are running in Safari? I tried to run Chrome Experiments in a Monterey 12.5 VM and it worked OK on my Ventura host.
I was able to reproduce the screen tearing by scrolling Elon Musk's Twitter account repeatedly in Chrome, but no crashes, though. My configuration is:
- macOS 12.5 (
21G72
) - Apple M1 Pro (
MacBookPro18,3
) - Tart 0.19.0
YouTube is completely unusable at first (you can't see the website content), and then the Chrome tab suddenly crashes. Unfortunately, no VM crash was observed.
Is it the same issue as #84? Default CPU cores for Monterey VMs is 2 where in Ventura they changed it to 4 kind of acknowledging the performance issue.
Host: Monterey 12.5, 14" M1 Pro VM: Monterey 12.5, cpu 4 (i.e. this isn't #84)
I was scrolling a GitHub log in Safari. The tearing is easy to reproduce. Just scroll a lot, it will eventually go disco. The crash is harder I had to keep trying to interact for a while. Eventually it crashed and the whole VM quit. It does seem a lot harder to get it to crash though.
UPDATE: I haven't had much success reliably reproducing the crash. I've had it crash once this morning. But that's it out of about 30 min of really trying.
I can work around with Firefox as it's not using GPU acceleration.
I was previously using Chrome for unrelated reasons, but Chrome (and Brave) no longer work in the 0.19 for me (possibly earlier as well—my brew upgrade jumped a couple of versions). Every load (including Chrome's own settings) fails with an Error code: 5.
@ryanbooker have you tried 0.20.0
or later? I wonder if https://github.com/cirruslabs/tart/pull/182 helped since it used the latest Xcode for building.
Unfortunately 0.20.0
still has the tearing issue.
I ran some geekbench on my test M1 Mac Mini and it seems the performance is on parity. I created a Monterey 12.5 VM with the same amount of CPU/Memory as the host on Ventura.
Let's keep the issue open if someone else will have the same issue. But it's hard to pin point the regression in your case since we can't reproduce crashes and from the syntactic benchmarks there seems to be no regressions.
I was previously using Chrome for unrelated reasons, but Chrome (and Brave) no longer work in the 0.19 for me (possibly earlier as well—my brew upgrade jumped a couple of versions). Every load (including Chrome's own settings) fails with an Error code: 5.
We are experiencing the same issue: Chrome does not load any website, fails with error code 5. It's not the fault of Tart, as other hypervisors using virtualization.framework have the same issue.
What worked for us was to use: arch -x86_64 ./Google\ Chrome --disable-gpu
@testingbot have you had a chance to try Ventura VMs to see if the issue is with the Virtualization.Freamework
or maybe GPU drivers that were updated in Ventura.
It is the same problem (and workaround) on macOS Ventura (beta 22A5373b) VM unfortunately.
I have the same issue when recording videos in CI tests for iOS. The recorded video is laggy with a lot of tearing. The tearing only happens when the CI runs inside a VM. Ventura host, Ventura VM. The VM has a lot of CPU and memory, so doubt it's a resource constraint.
Closing since it doesn't seem like a Tart issue.