NEC Cluster cacau introduction

This platform serves the following purpose. It enables development and computation of parallel programs on the Intel Xeon processors with Intel EM64T Technology. The two major parallel programming standards MPI and OpenMP are supported. Please note that you must limit the execution time of your jobs during daytime to guarantee the short turn around times that are necessary for development.

Hardware and Architecture

The HWW Xeon EM64T cluster platform consists of one front node for interactive access (cacau.hww.de) and several nodes for execution of parallel programs. The cluster consists of 210 dual socket and 2 quad socket nodes with 3.2GHz/3.0GHz/2.4GHz Xeon EM64T CPU's + 2/8/16/128 GByte memory on the nodes and two 2way frontend node with 2 Xeon EM64T 3.2GHz CPU's + 6GByte memory. Additionally a RAID system with 8 TByte and a GPFS with 15 TByte is available. The local disks of each node (58 GByte) serves as scratch disks. 2 nodes are installed with 128GB memory and a fast local disk with 1.7TB.

Features:

• Cluster of 210 dual SMPs nodes NEC servers with 2/8/16/128 GByte memory
• Frontend node is a 2way NEC Express5800/120Rg-2 server with 6GByte memory
• Node-Node interconnect Voltaire Infiniband(Switch:ISR9288) Network + Gigabit Ethernet
• Disk 8 TByte home/shared scratch + 1.2 TByte local scratch + 15 TByte GPFS parallel Filesystem
• Batch system: Torque, Maui scheduler
• Operating System: Tao Linux release 1 (Mooch Update 2)
• NEC HPC Linux software packages
• Intel Compilers
• Voltaire MPI
• Switcher/Module

Peak Performance: 	        3.9 TFLOP/s
Cores/node: 	                4
Memory: 	                1 TB
Shared Disk: 	                24 TB 
Local Disks/node: 	        80 GB
Number of Nodes: 	        212
Node-node data transfer rate: 	10 Gbps(Full bisectional: 20Gbps) infiniband

Short overview of installed compute nodes

Type	memory	Freq	cores	Disk	PBS Queue	PBS properties	Interconnect	nodes	number
1	2GB	3.2 GHz	2	80GB	-	mem2gb	infiniband	noco001-075, noco109-204	172
2	8GB	3.0 GHz	4	160GB	workq	-	infiniband	noco075-106	32
3	8GB	3.2 GHz	2	80GB	-	mem8gb	infiniband	noco205-208	4
4	16GB	3.2 GHz	2	80GB	workq	mem16gb	infiniband	noco209-210	2
5	128GB	2.4 GHz	8	1.7TB	pp	-	GigE	pp2 - pp3	2

Access

The only way to access cacau.hww.de (frontend node of NEC Cluster) from outside HWW net is through ssh. Information on how to set up ssh can be found on our webserver at ssh.

Usage

The frontend node cacau.hww.de is intended as single point of access to the entire cluster. Here you can set your environment, move your data, edit and compile your programs and create batch scripts. Interactive usage like run your program which leads to a high load is NOT allowed on the frontend node cacau.hww.de. The compute nodes for running parallel jobs are available only through the Batch system (see Batch System chap. 2) installed on the frontend node cacau.hww.de!

HOME directories

All user HOME directories for every compute node of the cluster are located on the master node cacau.hww.de. The compute nodes have the HOME directories mounted via NFS. On every node of the cluster the path to your HOME is the same. The filesystem space on HOME is limited by a quota of 50MB! Please note the Filesystempolicy in sec. 1.4! Default startup files (.profile, .cshrc,...) for your environmental settings can be found in: /usr/local/skel Only the default .profiles and the commands module or switcher support the HWW cluster features like MPI, Compiler settings,...(see Program Development chap. 3 and Environment Settings chap. 1.3.3).

SCRATCH directories

Local scratch

When allocating nodes using the batch queuing system (Torque), the system creates your own scratch area on each of the allocated nodes. The path to this local scratch area is stored in the environment variable SCRDIR (echo $SCRDIR) in your batch job shell. After your batch jobs are finished, the $SCRDIR will be removed automatically.

Global scratch

Another scratch you can get are global space on shared filesystems. There are 2 globel shared filesystems available on cacau:

• default

It's a filesystem which is available via NFS on all cacau cluster nodes and on the cacau frontend system

• gpfs

IBM GPFS filesystem shared globaly on different HWW Clusters To use it on cacau compute nodes, you need to create a file named '.gpfs' in your HOME directory (touch $HOME/.gpfs). The GPFS filesystem need some of the compute nodes memory. If you are short in memory on those nodes and you didn't need this filesystem, then please delete $HOME/.gpfs. If no such file found in your HOME, then the GPFS modules will not be loaded on the compute nodes.

You are responsible to obtain it from the system. To get access to this global scratch filesystems you have to use the workspace mechanism 1.3.2.