Hi Salvatore On 18/09/2024 10:21, Salvatore Bonaccorso wrote: > Resending to the correct bug (#1082081) > > On Wed, Sep 18, 2024 at 10:19:28AM +0200, Salvatore Bonaccorso wrote: >> Hi David, hi Emanuel, >> >> On Wed, Sep 18, 2024 at 08:29:54AM +0200, David Prévot wrote: >>> Control: clone -1 -2 >>> Control: reopen -2 >>> Control: found -2 linux/6.1.106-3 >>> Control: retitle -2 Performance issues on VM (virtio) >>> >>> Hi, >>> >>> Le Mon, Sep 16, 2024 at 11:37:07AM +0000, Kocher Emanuel, Bedag a écrit : >>> […] >>>> We still see issues with linux/6.1.106-3 in combination with virtio. >>> Thanks for the feedback. >>> >>>> Network performance is down to ~4 Mb/s measured with iperf3 instead of >>>> ~8 Gb/s (with linux/6.1.99-1 respectively 6.1.0-23-amd64). It is >>>> reproducible and booting the previous kernel resolves the issue. >>> Our network team also noticed performance issues, roughly 20 to 30% >>> increase speed on packages download once we reverted to 6.1.0-23-amd64 >>> on a mirror we host. Not sure it’s exactly the same issue as the one >>> initially reported, hence the clone, but I agree there is still an issue >>> with the current stable kernel. >> Thanks for reporting back. I will try to reproduce it so we can test >> against newer 6.1.y kernels as well. Do you have reproducers available >> to show it?
Yes, we are able to reproduce it on Nutanix hypervisors (virtio modules). There are different effects, e.g. copying a 1G file from an NFS-Share to such a server results in input/output errors (after approx 50s) and the NFS share going stale for a while, whereas on a healthy server the process takes approx. 6s. We see a lot of kernel messages like the one below (size varying between 1368, 1380, 1398 and 1448): eth0: bad gso: type: 1, size: 1448 Tell me if you want me to test something. Regards, Emanuel
