Message slow to construct for small message

[ggggggggg]

See this thread for context. Consider the following code (requires PR #141). message takes about 140 ns, while Message takes 3400 ns. Something like message should be available or the performance of Message should be made comparable for small payloads.

using ZMQ, BenchmarkTools

function message(s)
  b = IOBuffer()
  write(b,s)
  m = Message(b.size)
  m[:]=b.data
  m
end

@benchmark message("1")
@benchmark Message("1")

[ggggggggg]

I wrote a version of message with only one copy, which is a bit faster than the previous one. Then I benchmarked all 3 version (my two, plus ZMQ.Message) vs length of string. ZMQ.Message appears to win for strings longer than roughly 250 bytes. But I think there may be something else wrong, when I run the benchmark, it takes much longer to complete the portion for Message than for the other two versions, despite the fact that the sum of minimum times across all benchmarks is lower. Nonetheless, here is the plot I generated, followed by the code I used.

using ZMQ, BenchmarkTools, PyPlot
function Base.setindex!(m::Message, v, i::Integer)
    @boundscheck if i < 1 || i > length(m)
        throw(BoundsError())
    end
    unsafe_store!(pointer(m), v, i)
end
function message(x)
  m = Message(sizeof(x))
  mb = Base.AbstractIOBuffer(m,true,true,true,false,length(m))
  write(mb,x)
  m
end

function message_cp(s)
  b = IOBuffer()
  write(b,s)
  m = Message(b.size)
  m[:]=b.data
  m
end

function benchmark_length(f,l)
  s = randstring(l)
  b=@benchmark $f($s) evals=10
  minimum(b.times)
end
lengths = 2.^(0:24)
message_mtimes = [benchmark_length(message,l) for l in lengths]
message_cp_mtimes = [benchmark_length(message_cp,l) for l in lengths]
Message_mtimes = [benchmark_length(Message,l) for l in lengths]

figure()
plot(lengths, message_mtimes, label="message")
plot(lengths, message_cp_mtimes, label="message_cp")
plot(lengths, Message_mtimes, label="Message")
xlabel("bytes")
ylabel("minimum time to construct (ns)")
legend(loc="best")
xscale("log")
yscale("log")

[JamesWrigley]

I'm closing this because we now have the internal _MessageRef type that makes a copy and is used by default in Sockets.send():

julia> using ZMQ, BenchmarkTools

julia> @benchmark Message("1")
BenchmarkTools.Trial: 7228 samples with 200 evaluations.
 Range (min … max):  625.230 ns …  6.416 ms  ┊ GC (min … max):  0.00% … 86.68%
 Time  (median):       1.019 μs              ┊ GC (median):     0.00%
 Time  (mean ± σ):     3.423 μs ± 97.551 μs  ┊ GC (mean ± σ):  44.81% ±  1.65%

  ▁█▇▅▄▃▃▂▂▁▁                                                  ▁
  █████████████▆▆▆▆▆▆▆▆▆▄▆▆▄▆▄▄▄▄▄▆▅▄▅▄▅▅▄▃▅▅▄▅▄▁▄▃▃▄▃▄▁▃▃▄▃▃▄ █
  625 ns        Histogram: log(frequency) by time      12.1 μs <

 Memory estimate: 704 bytes, allocs estimate: 10.

julia> @benchmark ZMQ._MessageRef("1")
BenchmarkTools.Trial: 10000 samples with 997 evaluations.
 Range (min … max):  18.201 ns …  10.954 μs  ┊ GC (min … max):  0.00% … 99.47%
 Time  (median):     20.397 ns               ┊ GC (median):     0.00%
 Time  (mean ± σ):   30.878 ns ± 152.317 ns  ┊ GC (mean ± σ):  15.03% ±  4.67%

  ▂▅▅▇█▅▁ ▁▃                               ▂▄▄▄▄▄▂▂▁▂▂▂▃▂▁▁    ▂
  ███████████▆▅▅▆▆▄▃▄▃▄▃▄▃▁▄▁▁▁▄▃▄▁▁▃▃▄▄▃▁▃█████████████████▇▇ █
  18.2 ns       Histogram: log(frequency) by time      47.1 ns <

 Memory estimate: 80 bytes, allocs estimate: 1.