At the OpenMx meeting on 12/9/11 we discussed the following:
- The status of the grant proposal. Our strategy is to attempt to address any possible concerns or criticisms a reviewer may have. In particular liability concerns were discussed. The group believes that the grant can be phrased with in such a way that what we our proposing is seen in the same light (liability wise) as existing medical equipment such as MRI machines, etc. Framing goals for the grant in terms of a decision tree to address possible issues that become intractable was also discussed.
A test case which require a significant amount of computation time in the front of OpenMx have been provided by Michael Neale. These test cases contain a model with ~80 submodels. Each submodel does FIML and then a summation across all submodels is performed. The time for the test case has been reduced from 5 mins to 1 min. Several strategies for further reducing the computation time have been discussed and more test cases will be provided.
- A weird behavior of the implementation of parallel ordinal FIML has been discovered. For at least one model under several different architectures there is no speedup provided (and in some cases slowdown is exhibited) for several different numbers of cores. While several different theories about the cause more diagnostic data is needed. The group plans to gather this data in the upcoming week.
- The feature request for to supply the gradients of the target function and constraints, and the approximate hessian was discussed. It was decided that the group would solicit more feedback from users before deciding on an implementation strategy. More information on the feature request and the opportunity to solicit feedback can be found here.
- Progress on the C implementation of Rnpsol was reported. Strategies to iteratively test the implementation with the current build system was discussed.
- So far the website is reporting 11 downloads for the beta release of OpenMx 1.2. An email to SEMNET will be sent to try and encourage more beta testers. The existing list of OpenMx Beta testers will also contacted. The new transparent parallelization will be highly advertised in both messages.
- Possible issues on various different shared architectures running pbs related to the new OpenMx parallelization was discussed. Due to the lack of standards between different architectures the only solution is for users to set environment variables so that pbs can justly use the appropriate number of cores.
- The specification of Jacobian matrices for each constraint for gradients was discussed. Ultimately, the group decided to solicit feedback from potential users before choosing a specification.