Fig. 9

Effects of distribution of GPUs over nodes. Performance scaling of SIRT routines over a range of GPU counts on a cluster of 8 hosts with 4 GPUs each, with two different ways of distributing the GPUs over hosts: with “max GPUs/host” we cluster the GPUs as much as possible on hosts, while with “min GPUs/host” we use as few GPUs per host as possible. The time “without communication” is the time spent on actual computation, with communication between nodes disabled. The volume size is \(N^3\), the detector size is \(N^2\), and N projections are used. The time shown is the time for a single SIRT iteration