Snowflake is slow!

[jeehoonkang]

The snowflake branch is slower than master even when HP is not playing any role at all:

 name                     control ns/iter  variable ns/iter  diff ns/iter   diff %  speedup 
 multi_alloc_defer_free   2,060,746        2,293,428              232,682   11.29%   x 0.90 
 multi_defer              1,329,516        1,631,245              301,729   22.69%   x 0.82 
 multi_flush              12,558,094       32,184,340          19,626,246  156.28%   x 0.39 
 multi_pin                4,283,460        4,460,518              177,058    4.13%   x 0.96 
 single_alloc_defer_free  34               40                           6   17.65%   x 0.85 
 single_defer             17               28                          11   64.71%   x 0.61 
 single_flush             110              724                        614  558.18%   x 0.15 
 single_pin               7                8                            1   14.29%   x 0.88

I fully expect that flush() becomes slower, but not pin() or defer(). Should investigate the performance degradation using flamegraphs.

[jeehoonkang]

After manual inspection of the generated x86 assemblies, I found out that the introduction of Garbage enum negatively affects the performance quite severely. In particular, I could recognize two reasons of performance degradation: (1) Garbage was bigger than Deferred: 40 bytes vs. 32 bytes in x86-64; and (2) case discrimination was quite expensive.

I mitigated (1) by changing the size of Deferred to 24 bytes so that Garbage is now 32 bytes. Anyway Guard::defer() will be used quite rarely. Though I'm not sure how to solve (2). After changing the size of Deferred, the benchmark result for defer() and pin() becomes slightly better:

 name                     control ns/iter  variable ns/iter  diff ns/iter   diff %  speedup
 multi_alloc_defer_free   2,060,746        2,177,730              116,984    5.68%   x 0.95                                                                                                                 
 multi_defer              1,329,516        1,451,704              122,188    9.19%   x 0.92                                                                                                                 
 multi_flush              12,558,094       28,366,437          15,808,343  125.88%   x 0.44                                                                                                                 
 multi_pin                4,283,460        4,174,085             -109,375   -2.55%   x 1.03                                                                                                                 
 single_alloc_defer_free  34               36                           2    5.88%   x 0.94                                                                                                                 
 single_defer             17               24                           7   41.18%   x 0.71
 single_flush             110              608                        498  452.73%   x 0.18
 single_pin               7                8                            1   14.29%   x 0.88

I'm relatively happy with the performance of the current implementation...

[jeehoonkang]

The new benchmark result for a recent version of Crossbeam + Snowflake (9591f5e):

 name                     control ns/iter  variable ns/iter  diff ns/iter   diff %  speedup 
 multi_alloc_defer_free   4,142,880        4,459,418              316,538    7.64%   x 0.93 
 multi_defer              1,411,127        1,970,390              559,263   39.63%   x 0.72 
 multi_flush              12,473,802       14,810,773           2,336,971   18.74%   x 0.84 
 multi_pin                4,257,543        7,721,718            3,464,175   81.37%   x 0.55 
 single_alloc_defer_free  49               65                          16   32.65%   x 0.75 
 single_defer             17               36                          19  111.76%   x 0.47 
 single_flush             164              175                         11    6.71%   x 0.94 
 single_pin               7                16                           9  128.57%   x 0.44