Hardware
HPCC
hardware is located in
two data centers. The
public production
server, grendel, and
its file servers are in
the Experimenatl Sciences Building, TTU data center on
campus. All other
servers, the
antaeus (grid) cluster and its
common file system, are
in the Reese data
center. "Public" nodes
are available for any
TTU researcher.
"Private" nodes are
owned by individual
researchers and
administered by HPCC.
Antaeus private nodes
are available for public
non-priority use , and
grendel private nodes
are available for
non-priority use by the
group of owners of
private nodes.
Campus
The Grendel cluster is a node based system.
1. Compute Nodes: It has 210 dual slot quad core nodes with Intel E5405
processors for a total of 1680 3.0 GHz cores. Each node contains two Intel 5405
Quad Core 64 bit processors on a single board, as an SMP unit. The core frequency is
3.0GHz. Each node contains 16GB of memory.
2. The cores are connected with DDR Infiniband, with a rating of 20
peak teraflops and 16.5 HPL teraflops
3. Interconnect: 4X DDR Non-Blocking InfiniBand
4. A 70 terabyte Lustre file system runs over Infiniband
Reese
The
Reese systems have 6TB
of public and 12TB of
private shared Lustre
storage attached to two
clusters and several
grid servers.
The
antaeus cluster is a
mixture of public and
private nodes, connected
by Gigabit Ethernet.
1.
Private/public dual core
(192 cpu)
48
nodes (16 public) with
dual dual-core Intel
5160 processors (3.0
GHz) and 4GB of memory
each. Designated
compute-10-x,compute-11-1
through compute-11-16.
2.
Private quad core (64
cpu)
8
nodes with dual
quad-core Intel 5345
processors (2.33 GHz)
and 12GB of memory each.
Designated compute-11-17
through compute-11-24.
Performance
The
performance of one and
two nodes was measured
with
http://www.netlib.org/benchmark/hpl
HPLinpack and
http://www.tacc.utexas.edu/resources/software/#blas
Goto Blas. Usually the
best performance for HPL
is with ncpu/2 MPI
threads and 2 OpenMP
threads. The slower
network on antaeus is
reflected in the 2-node
scaling.
On
most codes, the
quad-cores are not as
much ahead of the
dual-cores as the HPL
performance appears to
show.
1.
Dual single-core 3.2 GHz
Irwindale: 10.7 GF (1
node), 19.9 GF (2 nodes
with Infiniband, 93%
scaling)
2.
Dual dual-core 3.0 GHz
5160: 32.5 GF (1 node),
45.4 GF (2 nodes with
Gigabit Ethernet, 70%
scaling)
