dgram: support multiple buffers

[mcollina]

This PR adds support for:

• socket.send(buf, port, host, [, callback])
• socket.send([buf1, buf2, ... ], port, host, [, callback])

Fixes#4302.

This is probably missing some benchmarks, and might like an optimization pass, but I would like to get some feedbacks first :).

cc @ronkorving@indutny@saghul

[indutny]

I guess arguments fit on the same line? ;)

[ronkorving]

Great start, thanks a ton :) What do the benchmarks say at this time?

[mcollina]

Current benchmarks comes out slightly slower on the "old" API. These are possibly due to: #4374 (diff)

before:
net/dgram.js len=1 num=100 type=send dur=5: 0.00120
net/dgram.js len=1 num=100 type=recv dur=5: 0.00020
net/dgram.js len=64 num=100 type=send dur=5: 0.07730
net/dgram.js len=64 num=100 type=recv dur=5: 0.01255
net/dgram.js len=256 num=100 type=send dur=5: 0.30837
net/dgram.js len=256 num=100 type=recv dur=5: 0.04958
net/dgram.js len=1024 num=100 type=send dur=5: 0.71537
net/dgram.js len=1024 num=100 type=recv dur=5: 0.12022

After:
net/dgram.js len=1 num=100 type=send dur=5: 0.00106
net/dgram.js len=1 num=100 type=recv dur=5: 0.00018
net/dgram.js len=64 num=100 type=send dur=5: 0.06785
net/dgram.js len=64 num=100 type=recv dur=5: 0.01109
net/dgram.js len=256 num=100 type=send dur=5: 0.27151
net/dgram.js len=256 num=100 type=recv dur=5: 0.04330
net/dgram.js len=1024 num=100 type=send dur=5: 0.70451
net/dgram.js len=1024 num=100 type=recv dur=5: 0.11787

I have no benchmarks yet on the new APIs.

[thefourtheye]

Use common.mustCall here, so that it will make sure that the test will fail if the callback is not invoked.

[mcollina]

@thefourtheye@indutny I have updated the tests to follow the current conventions. I will update also the other test to the same on another PR.

[saghul]

Not on this PR, but it would probably be nice to use uv_udp_try_send here, just in case we can send it on the spot.

[mcollina]

I would do that after this land. Do you see any issue with switching to uv_udp_try_send?
What is the difference anyway? on what conditions it can be sent on the spot?

[saghul]

It's not switching, the idea is to use it to try to send the datagram on the spot, because the socket might be writable. Have a look at how uv_try_write is used in stream_wrap.cc.

[mcollina]

I'm 👍 on that. I'll get it done after this one.

[mcollina]

I have done a pass of performance optimization in JS-land, which result in better performance for the single buffer use case (and offset/length is now on par with master).

Here are the full results:
https://gist.github.com/mcollina/e8210d26ca6ef2630e23

Unexpectedly, passing multiple buffers in an array is slower (benchmark/dgram/headers.js) than calling Buffer.concat() in js land for a large payload:

dgram/headers.js len=1024 num=100 chunks=1 headers=4 type=concat dur=5: 0.78476 dgram/headers.js len=1024 num=100 chunks=1 headers=4 type=multi dur=5: 0.67456 dgram/headers.js len=1024 num=100 chunks=2 headers=4 type=concat dur=5: 0.78631 dgram/headers.js len=1024 num=100 chunks=2 headers=4 type=multi dur=5: 0.63593 dgram/headers.js len=1024 num=100 chunks=4 headers=4 type=concat dur=5: 0.75965 dgram/headers.js len=1024 num=100 chunks=4 headers=4 type=multi dur=5: 0.63869 dgram/headers.js len=1024 num=100 chunks=8 headers=4 type=concat dur=5: 0.74356 dgram/headers.js len=1024 num=100 chunks=8 headers=4 type=multi dur=5: 0.59303 

The small buffer case (up to 256 bytes) is way faster than calling Buffer.concat().

Any ideas why this is happening? I expected it to be faster.

[thefourtheye]

@trevnorris

[trevnorris]

nit: function name is slightly incomplete? maybe something like fixAndSumBuffers(). this is nothing serious, and not a blocker.

[mcollina]

So happy this is released!! Thx to everyone!