PTRAM: A Parallel Topology-and Routing-Aware Mapping Framework for Large-Scale HPC Systems

Application-Driven Requirements for Node Resource Management in Next-Generation Systems

“…MPIPP [17] is a placement framework that takes into account the characteristics of the target hardware, but does not address current architectures with complex, hierarchical multicore nodes. Other work in this area include PTRAM [18], TopoMapping [19], RAHTM [20], TreeMatch [21], LibTopoMap [22], EagerMap [23], and Hier-TopoMap [24]. Vendor solutions tailored to MPI implementations include those by IBM [25], Cray [26], and HP [27].…”

Exploiting Geometric Partitioning in Task Mapping for Parallel Computers

“…Most previous non-contiguous approaches have represented tasks' communication patterns and network topologies as graphs; graph algorithms were then applied to find good mappings. Finding optimal topology mappings is NP-Complete [28], so heuristics are often used to reduce complexity (e.g., [7], [11], [12], [16], [17], [19], [22], [31], [32], [35]). We, instead, use inexpensive geometric partitioning to reorder tasks and processors based on their geometric locality, and use the reordering to map tasks that are "close" to each other geometrically to processors that are "close" to each other in the mesh or torus.…”

“…For example, application of MFZ to a 1D dataset is shown in Figure 5. With mapping of 1D tasks with MFZ ordering to 2D nodes using FZ as in Figure 3(d), the third-level cuts divide tasks (27,28), (3,11), (43, 35), and (51, 59). Each of these tasks are separated by only one hop along y in Figure 3(d).…”

“…The number of hops is reduced when there are torus wrap-around links; with wrap-around links, each message is sent using only a single hop. A node in the network, such as 39, is next to the nodes 7, 35,37,38,47,55, where each neighbor differs in a single bit as a Gray-code property. All of task 39's neighbors, 35, 37, 38 and 47, are placed in neighboring nodes.…”