Personal tools
You are here: Home Resources Computers Historical Computers at PDC Lenngren Hardware, Lenngren

Hardware, Lenngren

— filed under: , ,

Description of the hardware in Lenngren

Lenngren is a distributed memory computer form Dell consisting of 442 PowerEdge 1850 servers. Each node have two 3.4GHz "Nocona" Xeon processors and 8GB of main memory. The peak performance of the system is 6TFlop/s. The nodes are connected with gigabit ethernet for login and filesystem traffic. A high performance Infiniband network from Mellanox is used for MPI traffic.

Part of the machine is owned by two biochemistry groups at Stockholm University and KTH. There are just over 300 batch nodes available for general use.

Lenngren is named after Swedish 18th Century poet Anna Maria Lenngren.

Lenngren uses Intel Xeon processors with the EM64T extensions. EM64T stands for Extended Memory 64-bit Technology and is essentially a Intel-implementation of AMD's 64-bit extensions to the IA-32 (or x86) architecure called AMD64.

The major differences between EM64T/AMD64 and IA-32 are

  • A 64-bit wide flat address space is supported
  • The number of general purpose registers are increased from 8 to 16

    and their width is extended to 64 bits.

  • The number of XMM registers (used by SSE) are increased from 8 to 16.
  • The instruction set is expanded to take advantage of the new and wider registers.

The Xeon "Nocona" Processor

Peak 64 bit floating point performance

6.8 GFLOP/s @ 3.4GHz

L1I cache

12kB (trace cache)

L1D cache

16kB

L1D associativity

4 way

L1D cache line size

64 bytes

L1D access policy

Write-Through

L2 cache size

1MB

L2 cache associativity

8 way

L2 cache line size

64 bytes

Supports SSE, SSE2, SSE3 and the EM64T extensions

InfiniBand is a bit-serial switched network with a raw data rate of 250MB/s in each direction per serial link. The nodes in the cluster use "4X" links which are four serial links run in parallel giving a peak data rate of 1GB/s in each direction. The InfiniBand adapters are connected to the system through a PCI-Express X8 slot which has a theoretical bandwidth of 2GB/s. Nodes are interconnected through a switch fabric that consists of 27 24-port switches, each connected to 16 nodes and with four uplinks each to two 120-port backbone switches.

The Infiniband Network

Latency (one-way, measured)

4.8 us

Uni-directional peak bandwidth

1 Gbyte/s

Bi-directional peak bandwidth

2 Gbyte/s

The latency have been measured between two nodes that are on the same switch fabric.

Document Actions