Nuclear Physics Greenbook Presentation Astro,Theory, Expt

1
Nuclear Physics Greenbook Presentation(Astro,The
ory, Expt)

• Doug Olson, LBNL
• NUG Business Meeting
• 25 June 2004
• Berkeley

2
Reminding you what Nuclear Physics is.

• The mission of the Nuclear Physics (NP) program
  is to advance our knowledge of the properties and
  interactions of atomic nuclei and nuclear matter
  and the fundamental forces and particles of
  nature.
• The program seeks to understand how quarks bind
  together to form nucleons and nuclei, to create
  and study the quark-gluon plasma that is thought
  to have been the primordial state of the early
  universe, and to understand energy production and
  element synthesis in stars and stellar
  explosions.

3
(No Transcript)
4
Contents

• Questions
• Answers
• Some Observations

5
Questions about current needsand 3-4 years out

• 1. What are your most important processing needs?
    (e.g. single CPU speed, of parallel CPU's,
  memory, ...) 2. What are your most important
  storage needs?   (e.g. I/O bandwidth to disk,
  disk space, HPSS bandwidth, space, ...) 3. What
  are your most important network needs?   (e.g.
  wide-area bandwidth, bandwidth between NERSC
  resources, ...) 4. What are your most important
  remote access needs?   (e.g. remote login,
  remote visualization, data transfer,    single
  sign-on to automate work across multiple sites,
  ...) 5. What are your most important user
  services needs?   (e.g. general helpdesk
  questions, tutorials, debugging help, ...) 6. Do
  you have special software requirements, if so
  what? 7. Do you have special visualization
  requirements, if so what? 8. Is automating your
  work across multiple sites important?   Called
  distributed workflow.  Sites could be other large
    centers, a cluster, your desktop, etc. 9.
  Anything else important to your project?
• Asked to all PIs with NP awards
• 9 responders good cross section

6
Responses

• Astronomy
• Swesty, SUNY SB, TSI Collaboration
• Nugent, LBNL
• Nuclear Theory
• Ji, NCSU, QCD
• Pieper, ANL, Light nuclei
• Dean, ORNL, Nuclear many body problem
• Vary, Iowa State, Nuclear reactions
• Lee, Kentucky, Lattice QCD
• Experiment
• Klein, LBNL, IceCube
• Olson, LBNL, STAR

7
1. What are your most important processing needs?
  (e.g. single CPU speed, of parallel CPU's,
memory, ...)

• Theory
• Small cluster of nodes (1-16 nodes) and long run
  duration (12-24h or more)
• implementation of execution of single-processor
  tasks would be welcomed.
• Faster processors are always a good thing!
• Up to 2048 parallel (for SMMC)
• Need faster interprocessor b/w (for two other
  codes)
• Eq. 256 CPU Altrix is 7X faster than Seaborg
• total number of cycles obtained over a range of
  processors (50 to 500). generally .5 - 2
  Gbytes/processor are needed on processors with
  1-4x seaborg speed.
• Memory per CPU - by far the most important to our
  project (saves I/O to disk and/or cuts down on
  inter-node communication)

8
1. What are your most important processing needs?
  (e.g. single CPU speed, of parallel CPU's,
memory, ...)

• Astro
• Getting a little more CPU speed is ok, but BY FAR
  we need faster bandwidth between processors and
  between nodes.
• 1024-2048 processors now
• Lower latency communications
• Expt
• Not parallel algorithms, compute at PDSF other
  linux clusters

9
2. What are your most important storage needs?  
(e.g. I/O bandwidth to disk, disk space, HPSS
bandwidth, space, ...)

• Theory
• Bandwidth to disk and gpfs disk space (increase
  by 100X for disk space).
• Inode limit is a persistant pain, but can be
  lived with.
• HPSS bandwidth, space
• Single file system across machines (seaborg,
  newton)
• Astro
• Fine now, may change as we move more to 3-D.
• improved parallel I/O throughput to disk for gt
  1024 processor jobs
• increased scratch disk capacity
• Expt
• Database, MySQL in use now
• Disk - scalable size I/O performance, gt100TB, gt
  1 GB/sec
• HPSS size I/O
• Automated caching, replication I/O load
  balancing

10
3. What are your most important network needs?  
(e.g. wide-area bandwidth, bandwidth between
NERSC resources, ...)

• Theory
• Moving 20 GB datasets today ORNL-NERSC-MSU
• Moving 0.5 TB datasets in 2 years
  ORNL-NERSC-MSU-LLNL-PNNL
• bandwidth between NERSC resources
• Astro
• Improved throughput between NERSC, ORNL, and
  Stony Brook
• Expt
• WAN bandwidth end-to-end (means endpoints or
  other LAN effects are often the problem), labs
  universities

11
4. What are your most important remote access
needs?   (e.g. remote login, remote
visualization, data transfer,    single sign-on
to automate work across multiple sites, ...)

• Theory
• X-windowed system
• Data transfer is becoming an increasingly
  important need.
• ssh/scp with authorized keys is fine. one-time
  passwords would severely handicap my use of a
  local emacs and tramp to edit, view, and transfer
  files.
• Astro
• Single sign-on to allow process automation is
  very important right now. Of CRITICAL importance
  is avoidance of one-time authentication methods
  which would kill any hopes of scientific workflow
  automation.
• Some remote viz.
• Expt
• Data transfer
• Single sign-on across sites for automated workflow

12
5. What are your most important user services
needs?   (e.g. general helpdesk questions,
tutorials, debugging help, ...)

• Theory
• support/online-help for Windows-based X-servers
• Programming languages online-references or links
  to online-references would be great.
• General helpdesk and sometimes tutorials.
• Online tutorials (stored and indexed)
• Astro
• Biggest problems are dealing with new compiler
  bugs.
• Performance optimization, requires help people
  who have access to the IBM compiler group to code
  kernels tuned.
• Expt
• General user support and collaboration software
  installation is very good.
• Need troubleshooting across sites and WAN

13
6. Do you have special software requirements, if
so what?

• Theory
• Part of my plans involve solving a large sparse
  eigenvalue problem. Software like Aztec is going
  to be useful for this.
• Astro
• We continue to rely on the availability of HDF5
  v1.4.5 for our I/O needs on seaborg. HDF5 1.6.x
  will not suffice as we have uncovered
  show-stopping bugs in this release.
• Expt
• Community collaboration software (CERN, ROOT,
  )
• Current install/maintenance procedures work well

14
7. Do you have special visualization
requirements,if so what?

• Theory
• We would welcome introduction of visual debugging
  tools for fortran/C, especially for MPI or HPF
  programs, if possible of course.
• Astro
• Some, but most have been covered by the viz
  group.
• We continue to rely heavily on the NERSC viz
  group to help us address our viz needs.

15
8. Is automating your work across multiple sites
important?   Called distributed workflow.  Sites
could be other large   centers, a cluster, your
desktop, etc.

• Theory
• Yes. We are considering how to develop common
  component software for nuclear physics problems.
  The low-energy nuclear theory community will
  increasingly move towards integrated code
  environments. This includes data movement, and
  workflow across several sites. (We do this now
  with NERSC/ORNL/MSU).
• I do a fair amount of post processing with
  Speakeasy on my workstation. This involves mixing
  results from NERSC, Argonne's parallel machines,
  and Los Alamos' qmc at present.

16
8. Is automating your work across multiple sites
important?   Called distributed workflow.  Sites
could be other large   centers, a cluster, your
desktop, etc.

• Astro
• Yes! We are currently working with the SPA
  (Scientific Process Automation) team from the
  SciDAC Scientific Data Managment ISIC on
  automatic our workflow between NERSC and our home
  computing site at Stony Brook.
• Expt
• Yes. Experiment collaboration computing is spread
  across large small sites and desktops. Need
  more integration with security tools for a more
  seamless environment.

17
9. Anything else important to your project?

• Theory
• Nersc is a great help, keep up a great work!
• My biggest concern with NERSC at the present time
  is that it has fallen behind the curve on the
  machine front. While I still consider NERSC a
  valuable resource to my research, I have
  diversified significantly during this FY.
• Any performance tools, such as POE, that help
  diagnose the bottlenecks in a code and help 
  suggest routes to improvements.
• Astro
• Memory bandwidth and latency.
• Expt
• User management across site national
  boundaries. Separate user registration
  accounts across many sites will become too
  burdensome. Think single sign-on seamless!

18
Observations

• A strong need for greater inter-processor
  bandwidth
• Faster processors, more memory
• Single file system view across NERSC
• Greater parallel FS performance (gt1024)
• More space
• More/better data management tools
• Single sign-on across sites
• Help with Inter-site (WAN) issues
• Much scientific computing now has workflow across
  several sites