Fix benchmarks with `OsIpcReceiver` not being `Sync` anymore

[antrik]

The first commit is the minimal fix to get it working again.

The second patch switches to a different thread handling approach, which inexplicably produces lower overhead in some (unrelated) cases on my system...

[antrik]

This supersedes #121

(Well, at least I think it's a preferable fix...)

Would be great if someone with a more modern system (or maybe even a different OS?) could check whether the second patch produces a visible difference -- especially in the threaded case (i.e. 256 k and above); though a visible difference for the non-threaded case would also be relevant to know...

[vvuk]

I just tested with and without this -- the big data benchmark numbers definitely change, e.g. from:

test size_19_512k   ... bench:     595,488 ns/iter (+/- 64,696)
test size_20_1m     ... bench:   1,107,073 ns/iter (+/- 151,017)
test size_21_2m     ... bench:   1,976,821 ns/iter (+/- 180,263)
test size_22_4m     ... bench:   3,737,198 ns/iter (+/- 438,126)
test size_23_8m     ... bench:   7,096,525 ns/iter (+/- 887,525)

to

test size_19_512k   ... bench:     558,607 ns/iter (+/- 22,566)
test size_20_1m     ... bench:   1,196,693 ns/iter (+/- 93,622)
test size_21_2m     ... bench:   2,102,265 ns/iter (+/- 230,279)
test size_22_4m     ... bench:   3,826,037 ns/iter (+/- 212,121)
test size_23_8m     ... bench:   7,267,728 ns/iter (+/- 852,904)

But not enough to care I don't think. We should still land this, as without it it's impossible to benchmark.. if someone wants to have more precise benchmarking they can work on that afterwards.

[antrik]

@vvuk can you check whether they change between applying only the first patch or both? That's what I'm mostly concerned about... (You could also compare against #121 for another variant.)

[vvuk]

These are the latest Windows numbers, from my laptop (past thing was from a desktop). Second numbers look fine.

Before Sync change:

test create_channel ... bench:      14,904 ns/iter (+/- 1,757)
test size_00_1      ... bench:       1,671 ns/iter (+/- 136)
test size_01_2      ... bench:       1,667 ns/iter (+/- 137)
test size_02_4      ... bench:       1,665 ns/iter (+/- 186)
test size_03_8      ... bench:       1,664 ns/iter (+/- 165)
test size_04_16     ... bench:       1,708 ns/iter (+/- 211)
test size_05_32     ... bench:       1,643 ns/iter (+/- 114)
test size_06_64     ... bench:       1,681 ns/iter (+/- 166)
test size_07_128    ... bench:       1,691 ns/iter (+/- 143)
test size_08_256    ... bench:       1,715 ns/iter (+/- 164)
test size_09_512    ... bench:       1,842 ns/iter (+/- 237)
test size_10_1k     ... bench:       1,913 ns/iter (+/- 178)
test size_11_2k     ... bench:       2,065 ns/iter (+/- 108)
test size_12_4k     ... bench:       2,627 ns/iter (+/- 336)
test size_13_8k     ... bench:       3,155 ns/iter (+/- 235)
test size_14_16k    ... bench:      12,807 ns/iter (+/- 1,044)
test size_15_32k    ... bench:      16,779 ns/iter (+/- 2,277)
test size_16_64k    ... bench:      29,813 ns/iter (+/- 4,291)
test size_17_128k   ... bench:     338,392 ns/iter (+/- 105,054)
test size_18_256k   ... bench:     354,363 ns/iter (+/- 134,744)
test size_19_512k   ... bench:     463,017 ns/iter (+/- 179,602)
test size_20_1m     ... bench:     786,914 ns/iter (+/- 175,035)
test size_21_2m     ... bench:   1,254,104 ns/iter (+/- 470,615)
test size_22_4m     ... bench:   2,344,278 ns/iter (+/- 560,749)
test size_23_8m     ... bench:   4,597,099 ns/iter (+/- 739,417)

After first patch:

test create_channel ... bench:      14,603 ns/iter (+/- 6,303)
test size_00_1      ... bench:       1,639 ns/iter (+/- 186)
test size_01_2      ... bench:       1,639 ns/iter (+/- 150)
test size_02_4      ... bench:       1,633 ns/iter (+/- 165)
test size_03_8      ... bench:       1,633 ns/iter (+/- 185)
test size_04_16     ... bench:       1,601 ns/iter (+/- 135)
test size_05_32     ... bench:       1,580 ns/iter (+/- 164)
test size_06_64     ... bench:       1,650 ns/iter (+/- 204)
test size_07_128    ... bench:       1,626 ns/iter (+/- 119)
test size_08_256    ... bench:       1,676 ns/iter (+/- 161)
test size_09_512    ... bench:       1,795 ns/iter (+/- 190)
test size_10_1k     ... bench:       1,876 ns/iter (+/- 163)
test size_11_2k     ... bench:       2,023 ns/iter (+/- 188)
test size_12_4k     ... bench:       2,581 ns/iter (+/- 192)
test size_13_8k     ... bench:       3,277 ns/iter (+/- 251)
test size_14_16k    ... bench:      17,573 ns/iter (+/- 3,355)
test size_15_32k    ... bench:      18,197 ns/iter (+/- 3,273)
test size_16_64k    ... bench:      30,316 ns/iter (+/- 3,435)
test size_17_128k   ... bench:     315,150 ns/iter (+/- 37,126)
test size_18_256k   ... bench:     341,686 ns/iter (+/- 101,133)
test size_19_512k   ... bench:     444,938 ns/iter (+/- 117,830)
test size_20_1m     ... bench:     770,825 ns/iter (+/- 113,292)
test size_21_2m     ... bench:   1,243,743 ns/iter (+/- 185,992)
test size_22_4m     ... bench:   2,303,507 ns/iter (+/- 430,200)
test size_23_8m     ... bench:   4,457,642 ns/iter (+/- 533,357)

After second patch:

test create_channel ... bench:      15,004 ns/iter (+/- 1,947)
test size_00_1      ... bench:       1,642 ns/iter (+/- 149)
test size_01_2      ... bench:       1,642 ns/iter (+/- 135)
test size_02_4      ... bench:       1,640 ns/iter (+/- 130)
test size_03_8      ... bench:       1,638 ns/iter (+/- 242)
test size_04_16     ... bench:       1,628 ns/iter (+/- 117)
test size_05_32     ... bench:       1,526 ns/iter (+/- 600)
test size_06_64     ... bench:       1,631 ns/iter (+/- 140)
test size_07_128    ... bench:       1,552 ns/iter (+/- 464)
test size_08_256    ... bench:       1,673 ns/iter (+/- 164)
test size_09_512    ... bench:       1,788 ns/iter (+/- 186)
test size_10_1k     ... bench:       1,853 ns/iter (+/- 193)
test size_11_2k     ... bench:       2,016 ns/iter (+/- 207)
test size_12_4k     ... bench:       2,567 ns/iter (+/- 241)
test size_13_8k     ... bench:       3,259 ns/iter (+/- 235)
test size_14_16k    ... bench:      13,542 ns/iter (+/- 1,424)
test size_15_32k    ... bench:      17,592 ns/iter (+/- 2,304)
test size_16_64k    ... bench:      29,932 ns/iter (+/- 2,842)
test size_17_128k   ... bench:     276,133 ns/iter (+/- 53,773)
test size_18_256k   ... bench:     309,478 ns/iter (+/- 56,093)
test size_19_512k   ... bench:     381,911 ns/iter (+/- 98,733)
test size_20_1m     ... bench:     793,852 ns/iter (+/- 133,922)
test size_21_2m     ... bench:   1,216,504 ns/iter (+/- 251,077)
test size_22_4m     ... bench:   2,325,117 ns/iter (+/- 375,545)
test size_23_8m     ... bench:   4,535,233 ns/iter (+/- 764,413)

[antrik]

So just like on Linux, there is a lot of fluctuation in the threaded tests.

I suspect the lower numbers for 128k and 256k in the last run are also just a random fluctuation rather than actually better results?

At any rate, it doesn't look like either variant actually makes things worse :-)

[antrik]

I decided to drop the second patch, as it's really an unrelated change -- so this is just a trivial fix now.

[wafflespeanut]

@bors-servo r+

[bors-servo]

📌 Commit 44894df has been approved by Wafflespeanut

[bors-servo]

⌛ Testing commit 44894df with merge 1c12453...

[bors-servo]

☀️ Test successful - status-appveyor, status-travis