F# Akka.NET agents performance optimization when synchronizing - performance

Question

I am trying to solve the following problem. I have some agents running in real-time, with a large heartbeat of couple of milliseconds, and the order of operations they process for this reason are mostly deterministic (as the message processing is not the bottleneck).
Now, I am running a large amount of simulations of the system where I no longer have a heartbeat (otherwise it will take a couple of centuries) - but I need to make sure the order of operations are preserved. For this, I adopted the following solution: the simulator makes sure each agent has processed his message queue, by posting a dummy synchronization message and blocking while waiting for the answer. This does work for my application, but the time it takes is not intuitive - as a single threaded implementation would be an order of magnitude faster (I guess - x 100 ish - although I have not tested ).
I have isolated a small test that shows the issue, even trying to use another library, akka.net
type Greet =
| Greet of string
| Hello of AsyncReplyChannel<bool>
| Hello2
[<EntryPoint>]
let main argv =
let system = System.create "MySystem" <| Configuration.load()
let greeter = spawn system "greeter" <| fun mailbox ->
let rec loop() = actor {
let! msg = mailbox.Receive()
let sender = mailbox.Sender()
match msg with
| Greet who -> () // printf "Hello, %s!\n" who
| Hello2 -> sender.Tell(true)
| _ -> ()
return! loop()
}
loop()
let greeterF =
MailboxProcessor.Start
(fun inbox ->
async {
while true do
let! msg = inbox.Receive()
match msg with
| Greet who -> () // printf "Hello, %s!\n" who
| Hello reply -> reply.Reply true
| _ -> ()
}
)
let n = 1000000
let t1 = System.Diagnostics.Stopwatch()
t1.Start()
for i = 1 to n do
let rep = greeterF.PostAndReply(fun reply -> (Hello reply)) |> ignore
()
printfn "elapsed Mailbox:%A" t1.ElapsedMilliseconds
t1.Restart()
for i = 1 to n do
let res = greeter.Ask (Hello2)
let rr = res.Result
()
printfn "elapsed Akka:%A" t1.ElapsedMilliseconds
System.Console.ReadLine () |> ignore
0
Basically, both take about 10 seconds for a mere 1 millions of synchronizations - and not computation what-so-ever involved, and this is... unfortunate.
I am wondering if anybody has come across the same problem and if there is anyway to switch off the overhead forcing everything to run in a single threaded mode... something like better than de-activating all the cpus but 1 in the bios - or writing a clone of the whole system without the agents.
Any help well appreciated.

Answer 1

The reason for the Akka.NET version beeing slow here is how you communicate with the actor:
main process Task FutureActorRef !!ThreadPool!! greeter
Ask ---------------------->
Tell----------->
MailboxRun ----->
(greeter mailbox is empty) |
<--------------------------Tell
<--Complete task
<----------.Result
For each iteration, a TPL task will be created
Then a single message is sent to the greeter.
The main process then blocks while waiting for the response to come back.
The greeter replies back which in turn completes the task inside the FutureActorRef
Rinse and Repeat..
This design will cause Akka.NET to start and stop the greeters "mailbox run" for each message as the mailbox queue becomes empty for each iteration.
This results in threadpool schedulation for each single message that is passed.
It's a bit like entering your car, putting the pedal to the metal, then abruptly stop and step out of the car, and then repeating the procedure again.
That is just not a very effective way to travel fast.
#Aaronontheweb's suggestion will only have effect if you work out the above kinks in your code.
The mailbox needs to be able to constantly pick messages of the internal queue to work with messages in batches to achieve full throughput.
Instead, separate the producer from the consumer.
Create an actor that listens for the responses from your greeter.
And once that actor have processes your 1000000 messages, let that actor send a WorkCompleted message back to the consumer.
[Edit]
I gave it a shot myself, I don't know F# so it might not be completely idiomatic :)
open Akka
open Akka.Actor
open Akka.FSharp
type Greet =
| Greet of string
| Hello of AsyncReplyChannel<bool>
| Hello2
type Consume =
| Response
| SetSender
[<EntryPoint>]
let main argv =
let system = System.create "MySystem" <| Configuration.load()
let greeter = spawn system "greeter" <| fun mailbox ->
let rec loop() = actor {
let! msg = mailbox.Receive()
let sender = mailbox.Sender()
match msg with
| Greet who -> () // printf "Hello, %s!\n" who
| Hello2 -> sender.Tell(Response)
| _ -> ()
return! loop()
}
loop()
let consumer = spawn system "consumer" <| fun mailbox ->
let rec loop(count,sender : IActorRef) = actor {
if count = 1000000 then sender.Tell(true)
let! msg = mailbox.Receive()
match msg with
| Response -> return! loop(count+1,sender)
| SetSender -> return! loop(count,mailbox.Sender())
}
loop(0,null)
let n = 1000000
let t1 = System.Diagnostics.Stopwatch()
t1.Start()
for i = 1 to n do
greeter.Tell(Hello2,consumer)
let workdone = consumer.Ask SetSender
workdone.Wait()
printfn "elapsed Akka:%A" t1.ElapsedMilliseconds
System.Console.ReadLine () |> ignore
0
I updated your code to use a separate consumer for the actor responses and then reply back once all replies had been processed.
By doing so, your processing time is now down to 650ms on my machine.
If you want better throughput, you need to involve more actors to parallelize more.
I'm not sure if this helps in your specific scenario

Answer 2

Here is a slightly modified MailboxProcessor version:
module MBPAsync =
type Greet =
| Greet of string
| Hello of AsyncReplyChannel<bool>
let run n =
let timer = Stopwatch.StartNew ()
use greeter =
MailboxProcessor.Start <| fun inbox -> async {
while true do
let! msg = inbox.Receive()
match msg with
| Greet who -> () // printf "Hello, %s!\n" who
| Hello reply -> reply.Reply true
}
Async.RunSynchronously <| async {
for i = 1 to n do
do! Async.Ignore (greeter.PostAndAsyncReply Hello)
}
let elapsed = timer.Elapsed
printfn "%A" elapsed
The difference here is that this version uses PostAndAsyncReply and keeps the computation in an async workflow. On my quick test this seemed to be much faster that using PostAndReply, but YMMV.
The timings I get from the above MBP version look roughly like this:
> MBPAsync.run 1000000 ;;
00:00:02.6883486
val it : unit = ()
A comment earlier mentioned my Hopac library. Here is an optimized version using Hopac:
module Hop =
type Greet =
| Greet of string
| Hello of IVar<bool>
let run n =
let timer = Stopwatch.StartNew ()
let greeterCh = ch ()
do greeterCh >>= function
| Greet who -> Job.unit ()
| Hello reply -> reply <-= true
|> Job.forever
|> server
Job.forUpToIgnore 1 n <| fun _ ->
let reply = ivar ()
greeterCh <-- Hello reply >>.
reply
|> run
let elapsed = timer.Elapsed
printfn "%A" elapsed
The timings I get from the above Hopac version look roughly like this:
> Hop.run 1000000 ;;
00:00:00.1088768
val it : unit = ()

Answer 3

I'm not an F# developer, but I'm a core dev on Akka.NET. A couple of ideas for your scenario:
If you're only using a single actor for this work, you can try using a PinnedDispatcher - that way the actor runs on its own dedicated thread all the time. That will save you on unnecessary context switching overhead.
You can also set the throughput of the mailbox to be much higher for this PinnedDispatcher than the normal settings. i.e. set a throughput value of 10000 (or something) instead of the normal 25. Assuming that the contents of your mailbox grow in large bursts this should save you on mailbox synchronization overhead.
Here's what your dispatcher configuration might look like:
my-pinned-dispatcher {
type = PinnedDispatcher
throughput = 1000 #your mileage may vary
}
And then configure an actor to use it
C# Fluent Interface
var myActor = myActorSystem.ActorOf(Props.Create<FooActor>()
.WithDispatcher("my-pinned-dispatcher");
Config
akka.actor.deployment{
/greeter{
dispatcher = my-pinned-dispatcher
}
}
These are both options you can configure via HOCON in App.config or Web.config or you can use the fluent interface on the Props class to do this. Also worth noting: there's a bug with pinned dispatchers at the moment, but that should be fixed in our next maintenance release (v1.0.1,) which should be out next week.
Your mileage may vary, but this is what I would try - basically it's just designed to help reduce contention and overhead around a single actor.