select/cray: overload meaning of NodeAddr/NodeHostname in multiple-fronted mode
This implements Moe's suggestion for NodeAddr/NodeHostname semantics, NodeName "nid#####" (this is what SLURM will refer to the node as) NodeHostName "c0-0c0s0n1" (Cray's component ID, visible only with scontrol and sview's node display) NodeAddr "###" (hexadecimal X, Y and Z coordinates, visible only with scontrol and sview's node display) For example, palu> scontrol show node nid00189 NodeName=nid00189 Arch=XE CoresPerSocket=6 CPUAlloc=0 CPUErr=0 CPUTot=24 Features=(null) Gres=(null) NodeAddr=01E NodeHostName=c1-0c0s1n1 RealMemory=32000 Sockets=4 ... Please note: ~~~~~~~~~~~~ on XE systems each two nodes (0/1 and 2/3) on a node share the same network interface and hence are located at identical Y coordinates in the torus. To make tools such as smap work with these coordinates, we use "virtual" Y coordinates, computed as y_coord = 4 * cage + cpu; This scheme mirrors the one currently used to derive node coordinates on a SeaStar/XT system. 09_Cray-hostlist.diff
Please register or sign in to comment