** Next:** Field test cases
** Up:** parallel implementation
** Previous:** Parallel calculation of the

Our current bottleneck is a software design issue.
Much of SEPlib, assumes axes no larger than 2^{31} by
using integers. If we exceed this number we get semi-random
errors. Another potential problem is the way
we implemented the eignvector calculation. If we scale
to many nodes, the communication time will dominate,
and we will significantly degraded performance.
We have so far avoided this problem by running on a
low-latency, high-bandwidth network.

** Next:** Field test cases
** Up:** parallel implementation
** Previous:** Parallel calculation of the
Stanford Exploration Project

4/5/2006