performance question collective vs individual dset write

Hi,
Anyone have a feeling of whether a collective write or individual
writes would be faster in the following scenario: file etc. opened
collectively for MPI-I/O, being written to a lustreFS. 2%-10% of the
MPI ranks need to write hyperslabs to the dset. Typical number of MPI
ranks is ~2000. From googling around i get the sense that collective
IO is usually faster because it's coordinated, but given the low % of
MPI ranks needing to write to the data set should I go with individual
dset writes? (The two approaches are outlined here:
http://www.hdfgroup.org/hdf5-quest.html#par-nodata)

Thanks,
Izaak Beekman

ยทยทยท

===================================
(301)244-9367
UMD-CP Visiting Graduate Student
Aerospace Engineering
ibeekman@umiacs.umd.edu
ibeekman@umd.edu