Louhi User's Guide, the 2nd Edition > Louhi hardware > Cray XT system
Tehdyt toimenpiteet

Cray XT system

This chapter describes the structure of the mixed Cray XT4/XT5 system Louhi.

Louhi - Cray XT4 part

The XT4 part has 11 cabinets that include mostly 24 blades per cabinet. A blade includes four compute nodes. Each compute node has a single AMD Opteron 2.3 GHz quad-core processor. Each compute node has at least 4 GB memory (1 GB/core) and 96 nodes (1 cabinet) has 8 GB memory (2GB/core). The Cray SeaStar interconnect system directly connects all compute nodes in a 3D torus topology using Hyper-Transport links of Opteron processors. So 4 cores (one quad-core processor) share a same interconnect link.

Louhi - Cray XT5 part

The XT5 part has 9 cabinets and each of them has 24 blades per cabinet. As in XT4 a blade includes four compute nodes. Now each compute node has two AMD Opteron 2.3 GHz quad-core processors. Each compute node has at least 8 GB memory (1 GB/core) and 96 nodes (1 cabinet) has 16 GB memory (2GB/core). The interconnecting system is the same as in the XT4 part, one SeaStar interconnect per compute node. It means that 8 cores (two quad-core processor) share a same interconnect link.

Service nodes

In addition to compute nodes, Louhi has also 22 service nodes located to two XT4 cabinets (c0-0 and c0-1), running SuSE Enterprise Linux, used for login, I/O, boot and other service usage. Louhi has 70 TB of local fast

The Cray SeaStar2 interconnect system

The heart of the performance for massively parallel runs lies in the interconnection network between the processors. In Louhi, the interconnect is Cray SeaStar2 communication system. The Cray SeaStar2 interconnect system directly connects all compute nodes in a 3D torus topology using Hyper-Transport links of Opteron processors, providing a bandwidth-rich environment. The SeaStar2 interconnect system carries all message passing traffic (MPI) as well as all I/O traffic on both (XT4 and XT5) systems.

Cabinets

There are three critical differrencies between compute nodes:  the node is of the XT4 or XT5 type, and the memory per core (and thus per node), and the clock cycle of the processor is 2.3 or 2.7 GHz (PRACE nodes). The different kind of compute nodes should not be mixed when jobs are sent to execution. The following table shows which cabinets contains which kind of nodes (PRACE cabinets also included):

Node ID (NID) Cabinet Memory bus Node type Memory/core Memory/node Other
     24-95  c0-0  800 MHz  XT4  1 GB   4 GB  + Service Nodes
   148-223  c0-1  800 MHz  XT4  1 GB   4 GB  + Service Nodes
   256-351  c1-0  800 MHz  XT4  1 GB   4 GB  
   384-479  c1-1  800 MHz  XT5  1 GB   8 GB  
   512-607  c2-0  800 MHz  XT4  1 GB   4 GB  
   640-735  c2-1  800 MHz  XT4  2 GB   8 GB  
   768-863  c3-0  800 MHz  XT5  1 GB   8 GB  
   896-991  c3-1  800 MHz  XT5  1 GB   8 GB  
 1024-1119  c4-0  800 MHz  XT4  1 GB   4 GB  
 1152-1247  c4-1  800 MHz  XT4  1 GB   4 GB  
 1280-1375  c5-0  800 MHz  XT5  2 GB  16 GB  
 1408-1503  c5-1  800 MHz  XT5  1 GB   8 GB  
 1536-1631  c6-0  800 MHz  XT4  1 GB   4 GB  
 1664-1759  c6-1  800 MHz  XT4  1 GB   4 GB  
 1792-1887  c7-0  800 MHz  XT5  1 GB   8 GB  
 1920-2015  c7-1  800 MHz  XT5   1 GB   8 GB  
 2048-2143  c8-0  800 MHz  XT4  1 GB   4 GB  
 2176-2271  c8-1  800 MHz  XT4  1 GB   4 GB  
 2316-2399  c9-0  800 MHz  XT5  1 GB   8 GB  PRACE cabinet, + Service Nodes
 2432-2527  c9-1  800 MHz  XT5  1 GB   8 GB  PRACE cabinet

The 20 cabinets are physically on the floor of the computer hall in two rows, 10 cabinets in each row. In the column Cabinet there is the cabinet name, which is the first part of the node name. The second number in the cabinet name is the row number (0 or 1) and the first number is the cabinet number (0-8) in that row. The full node names are explained in Chapter Job launching command: aprun.