Optimize creation and serialization of closures

[amitmurthy]

Run

using Sockets, Distributed, Serialization

function srvr()
ls = listen(10000)
while true
    as = accept(ls)
    @async begin
        while true
            msg=deserialize(as)
            if isa(msg, Function)
                msg();
                serialize(as, ()->1)
            else
                serialize(as, msg)
            end
        end
    end
end
end

srvr()

in one process.

In a REPL

using Sockets, Distributed, Serialization

function testc_data(n)
    c = connect("localhost", 10000)
    @time for i in 1:n
        msg = serialize(c, 1)
        resp = deserialize(c)
        @assert resp == 1
    end
end

function testc_func(n)
    c = connect("localhost", 10000)
    @time for i in 1:n
        msg = serialize(c, ()->1)
        resp = deserialize(c)
        @assert resp() == 1
    end
end

Results

julia> testc_data(1000)
  40.453 milliseconds (22002 allocations: 1298 KB, 4.31% gc time)

julia> testc_func(1000)
 743.730 milliseconds (734 k allocations: 24448 KB, 0.19% gc time)

julia> testc_data(1000)
  36.762 milliseconds (22001 allocations: 1298 KB)

julia> testc_func(1000)
 741.483 milliseconds (741 k allocations: 24614 KB, 0.19% gc time)

Since the underlying model is function execution, as opposed to data messaging, just optimizing this will speed up parallel code that does a lot of remote calls.

@carnaval , originally the server code was not within a function. It is still slower but not as bad.

[vtjnash]

marking this as priority, because serialization is critical to #8745

[malmaud]

Is someone actively working on this? I might take a stab at it today otherwise.

[JeffBezanson]

But doesn't #8745 use a different serializer? Also the performance of serialization itself is not important there.

@malmaud Cool, go ahead!

[vtjnash]

it uses both

[ViralBShah]

testc_data(1000) works, but seems 10x slower (although this is on a mac and I am sure the reported numbers were on linux). testc_func(1) just hangs.