Talk: Blue Gene Information Session. Prentice Bisbal.

  1. Presented by Rutgers Discovery Informatics Institute
  2. This is not about how to use it or run it – there will be a separate boot camp on that.  This is just a basic info session
  3. A main motivation for this is to help get funding – it seems access to a top 500 computer helps grants get accepted
  4. Some major HPC problems are:
    1. Gap between proc and mem speed keeps increasing
    2. Heat because of increasing transistor density and clock frequency
    3. Communication latency and bandwidth also haven’t kept up with procs
    4. For max scalability, computation and communication should be balanced (overpowered procs generate too much heat)
    5. Full featured OSes have much more than is necessary (don’t care about videocards, USB, task switching)
  5. Blue/Gene is designed to address this issue
    1. Distributed memory arch (message passing)
    2. Slower procs w/ dual floating point (powerpc)
    3. Proprietary high-speed low latency networks (a few actually, different topologies available depending on the type of job)
    4. System on chip design, each node is about the size of 2 business cards sitting entirely under a heatsink
    5. Proprietary minimalist single task OS
    6. Nodes are modest (4 core, 2 gb ram, 512 mb ram/core)
  6. Computers fit 32 into a draw, are **really** sandwiched together right on top of each other alternating facing in, drawers go into a rack.  32 drawers/rack
  7. Limitations:
    1. Compute nodes don’t access external environment (including disk) directly, must go through I/O nodes.  Cant interact with jobs, cant send signals to jobs
    2. Smallest job you can run is 256 cores, need at least 64 nodes
  8. At Rutgers (currently, planning to expand), there are
    1. 2 racks = 2048 nodes, 8192 cores
    2. 850 mhz
    3. 7 tb filesystem
  9. OS supports subset of ops as linux, but is not linux.  Most important/relevant linux ops are present.  Os is diskless, single-tasking
  10. When nodes are idle they are completely shut down
  11. I/O nodres run real linux (but also stripped down)
  12. Front end node (what end users access):
    1. Suse Enterprise Linux 10 – full linux.  Need to cross compute for nodes
    2. Has IBM compilers and libraries for engineering and scientific computing
    3. Load-Leveler Scheduler
  13. Types of applications that are suitable for blue gene:
    1. Highly Parallelizable
    2. IPC that scales up to several thousand procs (min of 256 procs)
    3. Fixed computational resources
    4. Limited I/O, database ops.  I/O is slow, as always
    5. Modest Ram reqs
    6. Need source code because it has to be compiled for blue gene (means most commercial soft is out)
  14. There is a boot camp Oct 16-17.  Need to know MPI, probably not relevant.  Must RSVP.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: