Language-ext: Akka.NET shows much better performance benchmark then Process

[jcoder58]

I ran the ping-pong benchmark in Akka.NET (http://getakka.net/) and, if the numbers can be believed, Akka.NET's performance is orders of magnitude better then Process for message sending. Both Akka.net and Language-ext compiled in release mode with Any CPU.

Akka.net Ping Pong:

Worker threads:         2047
OSVersion:              Microsoft Windows NT 6.2.9200.0
ProcessorCount:         8
ClockSpeed:             2801 MHZ
Actor Count:            16
Messages sent/received: 30000000  (3e7)

ActorBase    first start time: 7.32 ms
ReceiveActor first start time: 25.93 ms

            ActorBase                          ReceiveActor
Throughput, Msgs/sec, Start [ms], Total [ms],  Msgs/sec, Start [ms], Total [ms]
         1,  4863000,       4.94,    6172.97,   4745000,       4.55,    6327.55
         5, 13495000,       3.52,    2226.95,  13026000,       3.35,    2306.68
        10, 16816000,       3.46,    1787.97,  16438000,       3.19,    1829.12
        15, 18170000,       5.60,    1656.70,  18126000,       4.12,    1659.36
        20, 20188000,       4.24,    1490.51,  18315000,       3.22,    1641.94
        30, 20229000,       3.35,    1487.07,  18484000,       3.19,    1626.61
        40, 17964000,       3.68,    1674.60,  18226000,       3.32,    1649.93
        50, 19659000,       5.00,    1531.84,  18856000,       3.42,    1595.30
        60, 19815000,       4.63,    1519.11,  19417000,       3.07,    1549.01
        70, 18844000,       3.52,    1596.22,  19828000,       3.36,    1517.10
        80, 19893000,       5.09,    1513.93,  18808000,       3.81,    1599.36
        90, 19023000,       4.78,    1581.97,  18427000,       3.54,    1631.95
       100, 19181000,       4.78,    1568.93,  18575000,       3.37,    1619.33
       200, 19933000,       3.46,    1509.04,  19243000,       3.16,    1563.03

language-ext Throughput:

Sleeping for 20 seconds whilst it warms up
Sleeping for 19 seconds whilst it warms up
Sleeping for 18 seconds whilst it warms up
Sleeping for 17 seconds whilst it warms up
Sleeping for 16 seconds whilst it warms up
Sleeping for 15 seconds whilst it warms up
Sleeping for 14 seconds whilst it warms up
Sleeping for 13 seconds whilst it warms up
Sleeping for 12 seconds whilst it warms up
Sleeping for 11 seconds whilst it warms up
Sleeping for 10 seconds whilst it warms up
Sleeping for 9 seconds whilst it warms up
Sleeping for 8 seconds whilst it warms up
Sleeping for 7 seconds whilst it warms up
Sleeping for 6 seconds whilst it warms up
Sleeping for 5 seconds whilst it warms up
Sleeping for 4 seconds whilst it warms up
Sleeping for 3 seconds whilst it warms up
Sleeping for 2 seconds whilst it warms up
Sleeping for 1 seconds whilst it warms up
Warm up sent 949384 messages. Running for real now...
917442 messages sent in 20 seconds.
That's 45872.1 messages per second

[louthy]

@jcoder58 The Process library has so far been through zero optimisation passes, so it's not surprising it's currently lagging behind Akka. I have mostly (over the past year) been bringing features into the library as it's 'quick enough' and didn't want to prematurely optimise. I use it for a large project that will over the next 6 months or so probably need some more juice, so expect significant improvements, although Akka parity isn't the priority - the quality of the API and system as a whole is more important for me;

This is personal opinion, so take it with a pinch of salt, but I am really no fan of the very OO/Java-esque style of Akka and that was the main reason I the Process library. I am also no fan of its untyped leanings; I have seen in the past the benefits to strongly typing the bridges between disparate systems. For example the Process library checks whether the actor you're sending a message to can actually receive it, so you get errors at the point-of-use rather than via dead-letters (even when messaging remotely). Some of the additional work this library does will almost certainly mean it can never get to Akka speeds, but I don't see any reason why it couldn't get to 10x with some careful optimisation.

[jcoder58]

@louthy: I agree with you about the Akka design, which is why I'm looking at Process. What they've done is very impressive, but I want to move to a more functional style, without having to jump to F#. F# is a great language, but the C# tooling still wins for me.

Process and Core seem to be what I'm looking for, so as long as you're aware of the performance issue and have plans to improve it, I'll continue to look at it.

[louthy]

@jcoder58 Yep, sure is. It's definitely on the near term radar as I also have a very real need to scale it up, probably by the end of the year.

[louthy]

@jcoder58 Just a follow up from this. I have done some initial optimisation (the very low hanging fruit), and have seen a speed up of a factor of 32.73x. _It won't be available for a while_, but I thought you'd be interested to know that your example of 45,872 messages/second would be of the order of 1,501,390 messages/second.

I think this bodes well for future optimisations.

[louthy]

Project has moved to https://github.com/louthy/echo-process