Deepcomp 7000 million billion times the machine mixed structure of high-performance computing cluster architecture, by using part of the cluster of Intel Xeon processors with Intel Itanium2 processors fat node components, suitable for different application calculated. But it should be noted that, both part of the node program binary code is not compatible. The cluster section includes the authentication server, the front-end machine, management nodes, login nodes, start node, the I / O nodes, the backup node and the compute nodes.
Which compute nodes, including blade and thick node configuration is as follows:
1140 the ordinary (blade) node, each blade configuration two quad-coreXeon processor E5450, clocked at 3.00GHz, 32GB of memory;
Thickness of 38 nodes, IBM X3950M2 server each Thick node configuration 16 quad-core Xeon processor X7350-based, clocked at 2.93GHz, 512GB memory;
12 visualization nodes (rack server), 2 Intel Xeon E5450, 32GB memory, Nvidia 9800GTS graphics card;
Blade configuration and calculation of the other nodes in the same include two authentication server, two front-end machine, two management nodes, login nodes, 12 start-node, 120 IO nodes, two backup node;
Operating systems: Red Hat Enterprise Linux Server release 5.1, Linux kernel 2.6.18-53.el5.
Fat nodes using two sets of SGI Altix4700 down object code compatible Deepcomp 6800 (the 6800 environment Deepcomp the part of the application, MPI program can be run directly) using the Intel Itanium2 processor, its configuration is as follows:
System consists of
Two compute nodes, one head node 2 file service node
◆ Fat nodes (compute nodes), node NUMA architecture, configured 94-way blade module, a total of 188 processors, 376 CPU core, shared memory is about 2.5TB, mainly used for the batch job;
◆ Head node, Taiwan fat nodes to achieve through the logical partition for a user logs on, the compiler and debugger, and submit jobs. Four-socket blade head node configuration, a total of 16 CPU cores, 64GB memory.
◆ File Services node by SGI Altix450 through logical partition functions become CXFS file system, the user is not directly accessible. A total of 384 Intel Itanium2 dual-core processors
◆ Model 9140M, 1.66GHz clock speed, 666MHz bus
◆ L1 Cache: 16KB (Data) + 16KB (instructions) L2 Cache: 256KB (Data) + 1MB (instructions), L3 Cache: 18MB
◆ TDP 104W
The peak 5Tflops of the system floating-point
Total amount of memory 5TB
Interconnect network nodes NUMALink interconnect (bidirectional bandwidth of 6.4GB / s, delay 1us); Gigabit Ethernet interconnect between nodes for management, NFS file system; compute nodes using two 20Gbps Infiniband interaction Company.
Operating system uses the the SUSE Linux 10SP2, kernel version to 2.6.16.60-0.21.default;
Other:
System network: using Gigabit Ethernet and 20Gbps Infiniband interconnect between nodes;
Job management system, LSF HPC 7.0.
Global shared file system
User $ HOME directory: / home_soft / home using SNFS file system;
Software installation directory: / home_soft / soft SNFS file system;
Work: SNFS file system / datastore directory links;
/Luster: user workspace, use the Lustre file system;
Datastore / userdata directory: user important data storage space using SNFS file system, data is periodically migrate to the tape library. / datastore / workspace directory: the users temporary work directory SNFSfile system.
Fat node dedicated work space (only fat node visible):
Workspace SGI CXFS shared file system.
Thick the node dedicated work space (the thick node only visible): /gpfs1, / gpfs2, IBM GPFS file system.