| |
|
|
|
|
SGI Altix Technical Summary |
Production Date: March 1, 2005
[cobalt.ncsa.uiuc.edu]
[login-co.ncsa.teragrid.org]
| Machine Name |
Model |
Function |
Processor type |
# Procs |
L3 Cache
(per processor) |
Memory |
Internal Interconnect |
| co-login1 |
SGI Altix 3300 |
Interactive use |
1.3 GHz Itanium 2
(5.2 GFLOPS peak) |
12 |
6 Mbytes |
96 Gbytes |
SGI NUMAlink 3 |
| co-compute1 |
SGI Altix 3700 Bx2 |
batch jobs |
1.6 GHz Itanium 2
(6.4 GFLOPS peak) |
512 |
9 Mbytes |
1 Tbyte |
SGI NUMAlink 4
InfiniBand 4x (to Prisms) |
| co-compute2 |
3 Tbytes |
co-viz1 - co-viz8
(8 systems) |
SGI Prism |
batch jobs |
1.6 GHz Itanium 2
GPU: ATI FireGL X2 |
8 CPUs
4 GPUs |
9 Mbytes |
co-viz1 - co-viz2: 16 Gbytes
co-viz3 - co-viz8: 8 Gbytes |
SGI NUMAlink 4
InfiniBand 4x (between Prisms
and to compute nodes) |
Common Attributes
| COMPONENT |
DESCRIPTION |
| Hardware & Networking |
Architecture |
Distributed Shared Memory (ccNUMA) |
| Parallel Filesystem |
SGI CXFS
| Scratch: |
43 TB
|
| Projects: |
86 TB
|
|
| Software |
Operating System |
SGI ProPack 4.0.3
Kernel: 2.6.5
(as of September 25 2006) |
| Compilers |
Intel 9.0: Fortran77/90/95 C C++
GNU: Fortran77 C C++
|
| Programming Model |
OpenMP, Message Passing Interface (MPI) |
| Program Analysis Tools |
Totalview |
| Floating point format |
IEEE |
| Batch System |
Portable Batch System (PBS Pro) with
MOAB Scheduler |
| Policies/User Limits |
Home directory disk quota |
50 GBytes |
| Interactive scratch quota |
None currently |
| Maximum disk per job |
No limits currently |
| Charging Algorithm |
# SUs = ServiceLevel * Time
where
ServiceLevel = 1.0 for Normal Interactive
Normal Batch
Normal Dedicated Batch
= 1.25 for Industrial Batch
Industrial Dedicated Batch
Timeshared jobs:
Time = Total CPU hrs
Dedicated jobs:
Time = Total Wall Clock Hrs * Number of CPUs
on the host
|
The NCSA SGI Altix System (cobalt) is primarily intended to run applications
with moderate to high levels of parallelism (32-512 processors). Applications
running with a relatively small number of processors (ie. less than 16 CPUs), MPI
codes that perform well in a distributed cluster environment, and condor-type
computations are better suited to other HPC resources available at NCSA.
In particular, the cobalt is intended for:
- applications requiring large shared-memory (over 250GB)
- large scale simulations and hero runs
- high-end vizualization (via access to the Prism systems)
- codes with high inter-processor communications or inherent load
balancing challenges which perform better in an SMP environment
- large-scale interactive data analysis
- on-demand computing
|
|
|
|
|
