Network contention frequently dominates the run time of parallel
algorithms and limits scaling performance. Most previous studies
mitigate or eliminate contention by utilizing one of several ap-
proaches: communication-minimizing algorithms; hotspot-avoiding
routing schemes; topology-aware task mapping; or improving global
network properties, such as bisection bandwidth, edge-expansion,
partitioning, and network diameter. In practice, parallel jobs
often use only a fraction of a host system. How do processor al-
location policies affect contention within a partition? We uti-
lize edge-isoperimetric analysis of network graphs to determine
whether a network partition has optimal internal bisection. In-
creasing the bisection allows a more efficient use of the network
resources, decreasing or completely eliminating the link con-
tention. We first study torus networks and characterize partition
geometries that maximize internal bisection bandwidth. We examine
the allocation policies of Mira and JUQUEEN, the two largest pub-
licly-accessible Blue Gene/Q torus-based supercomputers. Our
analysis demonstrates that the bisection bandwidth of their cur-
rent partitions can often be improved by changing the partitions'
geometries. These can yield up to a X2 speedup for contention-
bound workloads. Benchmarking experiments validate the predic-
tions. Our analysis applies to allocation policies of other net-
works.
