Attached are flamegraphs from running vm.StartTerminaVM with qcow2 and raw disk images. I noticed that when extracting the squashfs or tar.xz rootfs, the CPU is usually at 15% or lower utilization. We appear to be blocked primarily on fsync() when appending data clusters.

Container download and setup is going to grow the disk a lot. Maybe we should preallocate a certain amount (1-2 GiB), increase our block size, or grow the disk by a few clusters instead of just one.

Results from my sample sizes of 1.
                        
vm.StartTerminaVM total runtime

       eve        kevin
raw    1m12s      1m16s
qcow2  2m53s      3m10s

vm.StartTerminaVM container download/extraction only
       eve        kevin
raw    38s        35s
qcow2  2m24s      2m03s


 

Deleted: flamegraph_raw.svg 505 KB

flamegraph_raw.svg 505 KB Download

Deleted: flamegraph_qcow2.svg 4.7 MB

flamegraph_qcow2.svg 4.7 MB Download