Suggesting two little performance improvements about rms

[piotrp88]

I compared survival versus rms 4.4-2 as the first batch. Than I compared survival against rms 4.4-2 after implementing the above suggestions as the second batch. I repeated the same test using a 200MB dataset.

Using the 64MB dataset, the modified rms library performed ~18.15% better than the original one and using the 200MB dataset it performed even better, ~20.60%.

These are the results of my tests (timings are averaged per each batch; there are 3 runs per batch and 2 batches per dataset, 24 total runs):

# ---------- 64MB dataset (3 runs per batch) -----------
# survival VS rms 4.4.2
survival: 44.583333
rms: 39.735333
Speed gain: 10.87%

# survival VS rms 4.4-2 (using suggested edits)
survival: 43.524000
rms: 32.525000
Speed gain: 25.27%



# ---------- 200MB dataset (3 runs per batch) -----------
# survival VS rms 4.4-2
survival: 172.125667
rms: 155.115000
Speed gain: 9.88%

# survival VS rms 4.4-2 (using suggested edits)
survival: 171.674667
rms: 123.157000
Speed gain: 28.26%

As dataset size increases, performance gain increases too.

[harrelfe]

On the first item, I don't quite agree that not storing the lp should imply that certain summary statistics should not be computed.

On the second suggestion I'd like to know the time overhead of names(large vector) <- NULL

Frank

[piotrp88]

For the first item it would probably be useful to add an argument to cph() like one of the followings:

• cph(..., compute.gini=T, compute.dxy=T)
• a list of discrimination indexes to be computed which defaults to all possible values: cph(..., discr.indexes=c("R2", "Dxy", "g", "gr")) so if one specifies cph(..., discr.indexes=NULL) then no discrimination indexes will be computed while cph(..., discr.indexes=c("R2")) only computes R2 index.

More generally it would be better to let the user have control over which computations have to be done when performance matters, especially for computations which involve vectors or matrixes of n rows.

For the second suggestion the time overhead of names(large vector) <- NULL on my machine is cetrainly trascurable:

> huge.vector <- rnorm(100000000, 1, 2)
> system.time(names(huge.vector) <- NULL)
   user  system elapsed 
  0.170   0.237   0.407 

but I'm running MRO. You should try using a standard R version.