Computational Chemistry Code on Rocks Clusters at UCSD

1
Computational Chemistry Code on Rocks Clusters at
UCSD

• Jerry Greenberg UCSD/SDSC

2

• Large Advances in Raw Speed
• Increased access to database technology
• Remote access to remote resources via the Grid
• New Computational Algorithms

P R O B L E M S I Z E
lt 1000
Data Explosion
lt 100
lt 30
Architecture
_________________________________
Programming Complexity

workstation
MPP
MTA
cluster computing
easy
moderate
difficult
3

4
Scientific Web Services Environment
Resource Layer
Middle Layer
GLOBUS
Execution Service
HPC Resource
Event Service
SOAP
Storage Service
Storage
Soap Server
User Interface Layer
Job Manager Service
SOAP
5
Cheimistry Software installed on SDSC Meteor
Cluster (I)

• GAMESS ab initio Quantum Mechanics. Parallel
  communication by sockets, mpi or shmem.
• AMBER Classical Dynamics of biomolecules.
  Parallel communication via mpi.
• Gaussian 98 ab initio Quantum Mechanics.
  Parallel communication via shmem or LINDA.
• APBS. Calculation of electrostatic interactions
  in biomolecules. Parallel communication via
  mpi/myrinet.

6
Software installed on SDSC Meteor Cluster (II)

• NWChem ab initio Quantum Mechanics and Classical
  Dynamics. Parallel communication via shared
  memory and mpi/myrinet or TCGMSG.
• CE (Combinatorial Extension). Protein homology
  calculations. Parallel communication via
  mpi/myrinet.
• Euler Genetic Sequencing. Sequential code.

7
Why Clusters?

• Cheap (no facilities infrastructure e.g CSC)
• Cheap (no specialized hardware, fast commodity
  processors)
• Cheap (lots of academic and systems freeware)
• Hardware easily scalable.
• Systems software easily scalable. (with Rocks)
• Linux OS makes maintaining and improving legacy
  codes (usually developed under some flavor of
  UNIX ) relatively easy.

8
Cluster Problems/Areas for Improvement

• Batch Queue system (i. e. PBS/Maui) does not
  clean up after itself (e.g. runaway processes).
• Myrinet issues
• Unreliability
• Unreliability compounded by lack of job cleanup.
• Performance improvement over ethernet not
  commensurate with maintenance requirements and
  downtime.
• Limited debugging facilities that do not macth
  that available on proprietary UNIX systems e.g.
  dbx.
• Glitches with gnu compilers.
• Access to nodes are not restricted.

9
Conclusion