... that is the question. I have been working on an algorithm which takes an array of vectors as input, and part of the algorithm repeatedly picks pairs of vectors and evaluates a function of these two vectors, which doesn't change over time. Looking at ways to optimize the algorithm, I thought this would be a good case for memoization: instead of recomputing the same function value over and over again, cache it lazily and hit the cache.
Before jumping to code, here is the gist of my question: the benefits I get from memoization depend on the number of vectors, which I think is inversely related to number of repeated calls, and in some circumstances memoization completely degrades performance. So is my situation inadequate for memoization? Am I doing something wrong, and are there smarter ways to optimize for my situation?
Here is a simplified test script, which is fairly close to the real thing:
open System
open System.Diagnostics
open System.Collections.Generic
let size = 10 // observations
let dim = 10 // features per observation
let runs = 10000000 // number of function calls
let rng = new Random()
let clock = new Stopwatch()
let data =
[| for i in 1 .. size ->
[ for j in 1 .. dim -> rng.NextDouble() ] |]
let testPairs = [| for i in 1 .. runs -> rng.Next(size), rng.Next(size) |]
let f v1 v2 = List.fold2 (fun acc x y -> acc + (x-y) * (x-y)) 0.0 v1 v2
printfn "Raw"
clock.Restart()
testPairs |> Array.averageBy (fun (i, j) -> f data.[i] data.[j]) |> printfn "Check: %f"
printfn "Raw: %i" clock.ElapsedMilliseconds
I create a list of random vectors (data), a random collection of indexes (testPairs), and run f on each of the pairs.
Here is the memoized version:
let memoized =
let cache = new Dictionary<(int*int),float>(HashIdentity.Structural)
fun key ->
match cache.TryGetValue(key) with
| true, v -> v
| false, _ ->
let v = f data.[fst key] data.[snd key]
cache.Add(key, v)
v
printfn "Memoized"
clock.Restart()
testPairs |> Array.averageBy (fun (i, j) -> memoized (i, j)) |> printfn "Check: %f"
printfn "Memoized: %i" clock.ElapsedMilliseconds
Here is what I am observing: * when size is small (10), memoization goes about twice as fast as the raw version, * when size is large (1000), memoization take 15x more time than raw version, * when f is costly, memoization improves things
My interpretation is that when the size is small, we have more repeat computations, and the cache pays off.
What surprised me was the huge performance hit for larger sizes, and I am not certain what is causing it. I know I could improve the dictionary access a bit, with a struct key for instance - but I didn't expect the "naive" version to behave so poorly.
So - is there something obviously wrong with what I am doing? Is memoization the wrong approach for my situation, and if yes, is there a better approach?