F# compiler keep dead objects alive

Question

I'm implementing some algorithms which works on large data (~250 MB - 1 GB). For this I needed a loop to do some benchmarking. However, in the process I learn that F# is doing some nasty things, which I hope some of you can clarify.

Here is my code (description of the problem is below):

open System

for i = 1 to 10 do
    Array2D.zeroCreate 10000 10000 |> ignore    
    printfn "%d" (GC.GetTotalMemory(true)) 

Array2D.zeroCreate 10000 10000 |> ignore
// should force a garbage collection, and GC.Collect() doesn't help either
printfn "%d" (GC.GetTotalMemory(true))
Array2D.zeroCreate 10000 10000 |> ignore    
printfn "%d" (GC.GetTotalMemory(true))
Array2D.zeroCreate 10000 10000 |> ignore    
printfn "%d" (GC.GetTotalMemory(true))
Array2D.zeroCreate 10000 10000 |> ignore    
printfn "%d" (GC.GetTotalMemory(true))

Console.ReadLine() |> ignore

Here the output will be like:

54000
54000
54000
54000
54000
54000
54000
54000
54000
54000
400000000
800000000
1200000000

Out of memory exception

So, in the loop F# discards the result, but when I'm not in the loop F# will keep references to "dead data" (I've looked in the IL, and apparently the class Program gets fields for this data). Why? And can I fix that?

This code is runned outside Visual Studio and in release mode.

Answer 1

The reason for this behavior is that the F# compiler behaves differently in the global scope than in local scope. A variable declared at global scope is turned into a static field. A module declaration is a static class with let declarations compiled as fields/properties/methods.

The simplest way to fix the problem is to write your code in a function:

let main () =    
  Array2D.zeroCreate 10000 10000 |> ignore    
  printfn "%d" (GC.GetTotalMemory(true))
  Array2D.zeroCreate 10000 10000 |> ignore    
  printfn "%d" (GC.GetTotalMemory(true))
  // (...)
  Console.ReadLine() |> ignore

main ()

... but why does the compiler declare fields when you're not using the value and just ignore it? This is quite interesting - the ignore function is a very simple function that is inlined when you use it. The declaration is let inline ignore _ = (). When inlining the function, the compiler declares some variables (to store the arguments of the function).

So, another way to fix this is to omit ignore and write:

Array2D.zeroCreate 10000 10000 
printfn "%d" (GC.GetTotalMemory(true))
Array2D.zeroCreate 10000 10000 
printfn "%d" (GC.GetTotalMemory(true))
// (...)

You'll get some compiler warnings, because the result of expression is not unit, but it will work. However, using some function and writing code in local scope is probably more reliable.