I was able to solve the problem by creating two runners per machine - one with high-load tag and another with normal tag. For high-load runner I've used limit=1 flag to prevent it from taking more than one job simultaneously. And of course you have to mark every resource-consuming job with high-load tag in order to make thing work.