MPI framework for parallelized warmups

Algorithms
11

From Bob’s thread on parallel performance It sounds like the goal here is to decrease the time to ‘convergence’. Could we do something like

  1. For K chains, do adaption for N iterations and sample for M iterations, saving the mass matrix
  2. Calculate ESS
  3. Rerun with supplied mass matrix for additional N adaption iterations and sample for another M iterations
  4. Goto 2 until some arbitrary number of adaption steps have happened.
  5. Take ESS and divide it by the number of M samples to get a normalized ESS
  6. Divide that by the number of warmups to get a Normalized ESS per adaption iterations.
NESS= \frac{ESS}{Samples} \\ Hurdle = \frac{NESS}{Adaption}

5 and 6 I’m just sort of throwing out there, but it’s nice because it at least sets some form of a lower and upper bound on the performance metric (max is (1/1) and min of 0 can be either poor ESS or adaption going to infinity). Though I’m not sure it’s that interpretable which is a shame. It also can’t handle 0 adaption steps (though we could just slap a 1 in the denominator)

Would wall time be a metric here? It could be nice as I’m not sure how much the overhead vs. any data we can share between chains without making copies would effect overall speed.

1 Like