Efficiency updates:
- As expected, calculating the last common ancestor on the fly was responsible for a significant (>50%) fraction of the computational effort, so it’s worth calculating this once for the tree and storing the information.
- Current profiling table:
% cumulative self self total
time seconds seconds calls us/call us/call name
57.14 0.08 0.08 1604986 0.05 0.05 calc_gamma
35.71 0.13 0.05 802493 0.06 0.16 calc_var
7.14 0.14 0.01 1517 6.59 6.59 alloc_tips
0.00 0.14 0.00 837384 0.00 0.00 node_age
0.00 0.14 0.00 34891 0.00 0.00 calc_mean
0.00 0.14 0.00 2025 0.00 0.00 get_lca
0.00 0.14 0.00 1517 0.00 92.29 calc_lik
0.00 0.14 0.00 1517 0.00 0.00 log_normal_lik
0.00 0.14 0.00 1 0.00 0.00 fit_model
0.00 0.14 0.00 1 0.00 0.00 multimin
calc_gamma doesn’t need to be called the n2 times in calc_var, but could be stored from the calc_mean computation, so it would be called no more than calc_mean is. This would probably speed up calc_var, since it’s a significant fraction of that time, and leave alloc_tips as the most intensive function. Like get_lca, alloc_tips could be performed once for the entire tree and stored in the tree structure. Several other optimizations are quite natural but should wait until these primary inefficiencies are dealt with. Likewise any parallelization will be best applied at higher level (bootstrapping, maybe MCMC steps).
Note that we’re getting 1517 calls to the likelihood function in 0.14 seconds; better than 10,000 likelihood calculations a second.