Re: Information request/Amazon EC2

Posted by Douglas Roberts-2 on
URL: http://friam.383.s1.nabble.com/Re-Information-request-Amazon-EC2-tp3475158p3480830.html

Jack,

It would be a fun project to move some already running largish distributed ABM from a standard Linux cluster over to EC2.

If only my company would pay me to play just on the fun projects...

--Doug

On Thu, Aug 20, 2009 at 12:47 PM, Jack K. Horner <[hidden email]> wrote:
At 09:00 AM 8/20/2009, Doug Roberts wrote:
Doug,

Whether a given parallel computing system performs well enough running a message-passing-oriented Agent Based Modeling (ABM) application depends on, among other things,

       1.  How the agents are distributed across the processing
          elements (pes, nominally one microprocessor per pe) of the
          system.  Computational-mesh-oriented (CMO) applications that use
          message-passing services are sufficiently analogous to
          ABM-oriented applications that we can use mesh performance
          data to help bound what ABM performance is likely to be,
          given an allocation of agents per pe.

          In particular, it is not uncommon for CMO
          applications using ~50 state variable per cell to allocate
          ~100,000 cells per pe; state updates in such a system are
          accomplished by message-passing (using OMP or MPI) among cells.

          100,000 cells per pe is an empirically derived "rule of thumb",
          but it is roughly invariant across modern production-class
          compute nodes and a wide spectrum of mesh-oriented applications.

          For optimal performance, the cells allocated to a pe should
          be the set of cells that communicate most frequently with
          each other.  Sometimes a user can characterize that set
          through a propagation-rate function defined in the
          problem space (e.g., the speed of sound in a
          medium, the speed at which a virus travels from one agent
          to another, the speed of chemical reactions in a
          biological network).  Sometimes we don't know anything about
          the communication/propagation dynamics, in which case
          "reading" a pile of steaming chicken entrails predicts
          performance about as well as anything else.

          By analogy, if there were no more than ~50 state variables
          per agent in an ABM application, an allocation of up to
          100,000 tightly-communicating agents per pe would provide
          usable performance on many production-class clusters today
          (a cluster of PlayStations is an exception to
          this rule  of thumb, BTW).

          Allocating one agent per pe would be a vast waste of
          compute power for all except trivial problem setups.

          All of the above is useful only if the user can control
          the allocation of agents to pes.  Most production-class
          clusters, including the EC2, provide such controls.

          Note that this problem has to be addressed by the
          *user* in *any* cluster.



       2.  If the computation/communication ratio has to be near 1
          to obtain tolerable time-to-solution, the
          performance of the message-passing services matters
          hugely.  MPI and OMP have been optimized on only a few
          commercially available systems.  (A home-brew
          multi-thousand-node Linux cluster, in contrast, is nowhere
          near optimal in this sense.  Optimizing the latter, as
          a few incorrigibly optimistic souls have discovered,
          amounts to redesigning much of Linux process-management.
         If bleeding-edge performance matters, there is no free lunch.)


Jack




Jack K. Horner
P. O. Box 266
Los Alamos, NM  87544-0266
Voice:   505-455-0381
Fax:     505-455-0382
email:   [hidden email]



============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org



--
Doug Roberts
[hidden email]
[hidden email]
505-455-7333 - Office
505-670-8195 - Cell

============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org