To prevent fragmentation, we currently only fill hosts up to a threshold (currently 72%). This is to avoid overloading hosts, especially since multiple VMs on the same host share some resources. However, this threshold causes inefficiencies, especially for hosts with fewer but more powerful cores like AX102. Since AX102 hosts have fewer but stronger cores, they are less affected by resource contention. This should improve premium runner availability, especially for higher vCPU requests like 30 vCPUs jobs, since we pack others more.
27 KiB
27 KiB