Parallel Gaussian Performance

["Ole W. Saastad" <ole[ AT ]scali.no>]

Parallel Gaussian
 Gaussian Inc. supplies a parallel version of g98 that
 uses Linda from Scientific Computing Associates.
 My questions are concerned about how well does the
 combination  of g98 and Linda perform? How well does
 it scale ? Do memory requirements also scale well ?
 What are the experience of the cluster users around ?
 I am interested in performance of Linda and Gaussian
 on linux clusters, both standard Beowulf clusters and
 tightly coupled clusters with Myrinet or Dolphin interconnect.
 How is the communication done with Linda ? The paradigm
 is "tuple space", but in reality some underlying layer
 must be present. How is this done ? Shared memory model
 or message passing ?
 I have seen (posted somewhere on the net) that there exist
 a MPI version of g94/g98. Have anyone experience with this
 version of g98(g94) ? A MPI version of Gaussian would not
 require the purchase of the Linda system.  For clusters with
 fast interconnect and fast MPI the MPI version of g98 is
 expected to yield higher performance than Linda/g98. How
 is this in practice ?
 Have anyone experience of running the MPI Gaussian on a cluster
 of Linux machines ?
 Again how is the performance on Beowulf, Myrinet or
 Dolphin interconnected systems ? How well does it scale ?
 When does it start to roll off (at 16 cpus or 32, 64 or even higher?).
 Best regards
 Ole W Saastad