I have to simulate a discrete environment in F#, to be called by Python, for a reinforcement learning problem. I had a function with primitive types (mainly float) to make the exchange of data smoother. Now I am in the position to run this function many times with different data, so to run it in parallel seems a good idea.
I have the following code:
type AscentStrategy = |Strategy of seq<float>
let simulateAscent env ascentLimiter initState (sequenceOfDepths:seq<float>) =
//let infinitSeqOfConstantValues = (fun _ -> constantDepth) |> Seq.initInfinite
sequenceOfDepths
|> Seq.scan ( fun ( nextState, rew, isTerminal, _ ) depth -> getNextEnvResponseAndBoundForNextAction(env, nextState , depth , ascentLimiter) ) ( initState, 0.0 , false, 0.0)
|> SeqExtension.takeWhileWithLast (fun (_ , _, isTerminalState, _) -> not isTerminalState)
|> Seq.toArray
and then
let simulateStrategy ({MaxPDCS = maxPDCS ; MaxSimTime = maximumSimulationTime ; PenaltyForExceedingRisk = penaltyForExceedingRisk ;
RewardForDelivering = rewardForDelivering ; PenaltyForExceedingTime = penaltyForExceedingTime ; IntegrationTime = integrationTime
ControlToIntegrationTimeRatio = controlToIntegrationTimeRatio; DescentRate = descentRate; MaximumDepth = maximumDepth ;
BottomTime = bottomTime ; LegDiscreteTime = legDiscreteTime } : SimulationParameters) (Strategy ascentStrategy : AscentStrategy) =
let env, initState , ascentLimiter , _ = getEnvInitStateAndAscentLimiter ( maxPDCS , maximumSimulationTime ,
penaltyForExceedingRisk , rewardForDelivering , penaltyForExceedingTime ,
integrationTime ,
controlToIntegrationTimeRatio,
descentRate ,
maximumDepth ,
bottomTime ,
legDiscreteTime )
ascentStrategy
|> simulateAscent env ascentLimiter initState
finally I call the function for testing:
let commonSimulationParameters = {MaxPDCS = 0.32 ; MaxSimTime = 2000.0 ; PenaltyForExceedingRisk = 1.0 ; RewardForDelivering = 10.0; PenaltyForExceedingTime = 0.5 ;
IntegrationTime = 0.1; ControlToIntegrationTimeRatio = 10; DescentRate = 60.0; MaximumDepth = 20.0 ; BottomTime = 10.0; LegDiscreteTime = 0.1}
printfn"insert number of elements"
let maxInputsString = Console.ReadLine()
let maxInputs = maxInputsString |> Double.Parse
let inputsStrategies = [|0.0 .. maxInputs|] |> Array.map (fun x -> Seq.initInfinite (fun _ -> x ) )
let testParallel = inputsStrategies
|> Array.Parallel.map (fun x -> (simulateStrategy commonSimulationParameters ( Strategy x )) )
I have compared this with Array.map and, while it is faster and uses 70% of the CPU on my laptop, still does not seem to use the whole processing power. I have run it on a machine with many more cores ( ~50) and it barely increases the CPU usage (it gets up to 3/4% of total usage with 50ish independent inputs). I think there must be a deadlock generated somewhere, but how can I detect and get rid of it?
Also, why does this happen? One of the advantages of functional programming, as I see it, is also to be able to parallelize easily.
PS: SeqExtension.takeWhileWithLast is a function I have found on SO, kindly provided by Tomas Petricek in one of his brilliant answers, if needed I can post it.
PPS: env is the environment, whose type is defined as:
type Environment<'S, 'A ,'I> = |Environment of (State<'S> -> Action<'A> -> EnvironmentOutput<'S ,'I>)
I have tried the same with Async.Parallel and ParallelSeq, reporting the same problem.
Would a message-based solution solve the problem>? I am looking into it, although I am not familiar at all, but would it be a good way of getting the code parallel, using MailboxProcessor?
Following my question, I have tried also this great library for parallel code, based on streams of data. https://nessos.github.io/Streams/.
I have added the following code:
let nessosResult = inputsStrategies
|> ParStream.ofArray
|> ParStream.map simulateStrategy
|> ParStream.toArray
I have defined an ad hoc type for inputStrategy (basic the old tuple I had) so that simulateStrategy accepts only one input. Unfortunately the problem seems very well hidden somewhere. I attach a graph with CPU usage. Time spent on my machine for the different cases is: ~8.8 sec (sequential); ~6.2 sec (Array.Parallel.map); ~ 6.1 sec (Nessos.Streams)
I have found that server garbage collection is necessary to get the best parallel performance on .NET. Something like this in your app.config:
<configuration>
<runtime>
<gcServer enabled="true" />
</runtime>
</configuration>