What does this equation mean?

A minimal model makes the difficulty explicit. If each of a fleet’s ranks is independently healthy and idle with probability p at a random instant, the probability that all n ranks a job needs are simultaneously available is $P_{\text{admit}}(n) \approx p^{\,n}$. This is a simplification — real fleets are not independent, and real schedulers do not wait for a spontaneous coincidence — but it explains why gang admission gets combinatorially harder, not linearly harder, as jobs grow, and why every production scheduler in this space is built around avoiding literal simultaneous-availability matching rather than performing it. Google’s Borg, one of the earliest cluster managers to operate at this scale, achieves its utilization through “admission control, efficient task-packing, over-commitment, and machine sharing,” and explicitly uses “scheduling policies that reduce the probability of correlated failures” [ 10 ] — packing and correlation-aware placement are both ways of making the effective p in that equation behave better than raw hardware uptime would suggest. Microsoft’s Singularity goes further and removes the need to solve admission as a fixed matching problem at all: “all jobs in Singularity are preemptable, migratable, and dynamically resizable (elastic) by default,” so a live job can be “transparently preempted and migrated to a different set of nodes, cluster, data center or a region and resumed exactly from the point where the execution was preempted” [ 11 ] . Elasticity turns a hard combinatorial admission problem into a softer one: the job does not need its final placement instantly, only a placement it can be moved out of without losing correctness.