runtime: use SwissTable #54766
Comments
gopherbot
added
the
compiler/runtime
Performance comparisonHere the SwissTable is just a basic version. To avoid a CL bringing too many changes, it does not use SIMD and is compatible with the previous hashmap memory layout(it means many optimizations are not included in this version). Its performance still has room for improvement; we will add these optimizations in subsequent CLs. In general, the performance changes for the basic version of SwissTable are as follows.
Platform
Benchmark-1Located in https://github.com/zhangyunhao116/gomapbench This benchmark contains performance comparisons of common operations of common types in different situations. INFOname old time/op new time/op delta MapIter/Int/6-16 60.2ns ± 1% 55.7ns ± 0% -7.35% (p=0.000 n=9+7) MapIter/Int/12-16 121ns ±10% 102ns ± 3% -15.51% (p=0.000 n=10+10) MapIter/Int/18-16 173ns ± 2% 165ns ± 1% -4.58% (p=0.000 n=10+8) MapIter/Int/24-16 214ns ± 4% 194ns ± 3% -9.46% (p=0.000 n=10+10) MapIter/Int/30-16 287ns ± 3% 272ns ± 2% -5.22% (p=0.000 n=10+9) MapIter/Int/64-16 609ns ± 0% 572ns ± 3% -6.10% (p=0.000 n=7+10) MapIter/Int/128-16 1.21µs ± 4% 1.17µs ± 2% -3.63% (p=0.004 n=10+10) MapIter/Int/256-16 2.46µs ± 2% 2.23µs ± 2% -9.31% (p=0.000 n=10+10) MapIter/Int/512-16 4.98µs ± 4% 4.38µs ± 2% -12.22% (p=0.000 n=10+10) MapIter/Int/1024-16 9.94µs ± 3% 8.84µs ± 2% -11.11% (p=0.000 n=10+10) MapIter/Int/2048-16 20.1µs ± 3% 17.8µs ± 1% -11.33% (p=0.000 n=10+10) MapIter/Int/4096-16 39.7µs ± 1% 35.5µs ± 2% -10.49% (p=0.000 n=10+10) MapIter/Int/8192-16 79.3µs ± 2% 71.8µs ± 1% -9.42% (p=0.000 n=10+10) MapIter/Int/65536-16 631µs ± 2% 571µs ± 1% -9.52% (p=0.000 n=10+9) MapAccessHit/Int64/6-16 4.24ns ± 1% 3.71ns ± 1% -12.37% (p=0.000 n=8+9) MapAccessHit/Int64/12-16 6.61ns ± 6% 6.63ns ±11% ~ (p=0.796 n=9+9) MapAccessHit/Int64/18-16 6.74ns ± 4% 7.01ns ± 1% +4.02% (p=0.002 n=9+8) MapAccessHit/Int64/24-16 7.81ns ± 8% 8.17ns ± 5% +4.62% (p=0.034 n=10+8) MapAccessHit/Int64/30-16 6.23ns ± 6% 6.99ns ± 1% +12.34% (p=0.000 n=10+7) MapAccessHit/Int64/64-16 6.68ns ± 6% 7.22ns ± 2% +8.08% (p=0.000 n=10+9) MapAccessHit/Int64/128-16 6.98ns ± 6% 7.32ns ± 6% +4.93% (p=0.004 n=10+10) MapAccessHit/Int64/256-16 7.12ns ± 2% 7.28ns ± 2% +2.21% (p=0.002 n=9+10) MapAccessHit/Int64/512-16 7.68ns ± 4% 7.40ns ± 2% -3.65% (p=0.000 n=9+10) MapAccessHit/Int64/1024-16 8.85ns ± 2% 7.35ns ± 2% -16.92% (p=0.000 n=10+9) MapAccessHit/Int64/2048-16 11.2ns ± 3% 7.7ns ± 1% -31.23% (p=0.000 n=10+8) MapAccessHit/Int64/4096-16 14.1ns ± 1% 7.9ns ± 1% -43.96% (p=0.000 n=7+8) MapAccessHit/Int64/8192-16 16.2ns ± 2% 8.2ns ± 2% -49.14% (p=0.000 n=10+10) MapAccessHit/Int64/65536-16 19.6ns ± 2% 10.5ns ± 1% -46.18% (p=0.000 n=10+9) MapAccessHit/Int32/6-16 3.78ns ± 1% 3.86ns ± 1% +2.12% (p=0.000 n=10+8) MapAccessHit/Int32/12-16 6.23ns ± 6% 6.16ns ± 4% ~ (p=0.483 n=10+9) MapAccessHit/Int32/18-16 6.42ns ± 5% 6.99ns ± 3% +8.91% (p=0.000 n=10+9) MapAccessHit/Int32/24-16 7.46ns ± 5% 8.60ns ± 8% +15.32% (p=0.000 n=9+10) MapAccessHit/Int32/30-16 6.07ns ± 5% 6.97ns ± 1% +14.88% (p=0.000 n=10+8) MapAccessHit/Int32/64-16 6.50ns ± 4% 7.18ns ± 5% +10.34% (p=0.000 n=9+10) MapAccessHit/Int32/128-16 6.68ns ± 6% 7.15ns ± 4% +7.12% (p=0.000 n=10+10) MapAccessHit/Int32/256-16 6.81ns ± 4% 7.32ns ± 4% +7.45% (p=0.000 n=10+10) MapAccessHit/Int32/512-16 7.40ns ± 3% 7.30ns ± 3% ~ (p=0.156 n=9+10) MapAccessHit/Int32/1024-16 8.46ns ± 2% 7.38ns ± 1% -12.79% (p=0.000 n=10+9) MapAccessHit/Int32/2048-16 11.0ns ± 3% 7.6ns ± 2% -31.46% (p=0.000 n=10+10) MapAccessHit/Int32/4096-16 13.9ns ± 2% 7.8ns ± 2% -44.12% (p=0.000 n=10+10) MapAccessHit/Int32/8192-16 15.8ns ± 2% 8.1ns ± 1% -48.99% (p=0.000 n=10+10) MapAccessHit/Int32/65536-16 19.0ns ± 2% 10.3ns ± 2% -45.85% (p=0.000 n=10+10) MapAccessHit/Str/6-16 13.5ns ± 1% 12.7ns ± 2% -5.88% (p=0.000 n=9+10) MapAccessHit/Str/12-16 9.77ns ±11% 9.40ns ± 2% ~ (p=0.460 n=10+8) MapAccessHit/Str/18-16 9.05ns ± 7% 9.47ns ± 1% +4.66% (p=0.001 n=9+9) MapAccessHit/Str/24-16 9.92ns ± 7% 10.07ns ± 9% ~ (p=0.604 n=9+10) MapAccessHit/Str/30-16 8.46ns ± 7% 9.43ns ± 1% +11.50% (p=0.000 n=10+8) MapAccessHit/Str/64-16 8.91ns ± 5% 9.57ns ± 2% +7.39% (p=0.000 n=10+8) MapAccessHit/Str/128-16 9.82ns ± 4% 10.58ns ± 5% +7.75% (p=0.000 n=10+10) MapAccessHit/Str/256-16 11.8ns ± 4% 11.4ns ± 1% -3.01% (p=0.008 n=10+8) MapAccessHit/Str/512-16 14.6ns ± 2% 11.9ns ± 2% -18.89% (p=0.000 n=8+9) MapAccessHit/Str/1024-16 17.9ns ± 2% 12.7ns ± 1% -29.02% (p=0.000 n=10+9) MapAccessHit/Str/2048-16 21.5ns ± 3% 13.0ns ± 2% -39.43% (p=0.000 n=10+9) MapAccessHit/Str/4096-16 25.1ns ± 1% 13.2ns ± 2% -47.63% (p=0.000 n=10+10) MapAccessHit/Str/8192-16 26.6ns ± 1% 14.2ns ± 1% -46.62% (p=0.000 n=9+9) MapAccessHit/Str/65536-16 31.8ns ± 1% 17.3ns ± 2% -45.64% (p=0.000 n=10+10) MapAccessMiss/Int64/6-16 8.03ns ± 1% 7.35ns ± 1% -8.39% (p=0.000 n=10+10) MapAccessMiss/Int64/12-16 10.2ns ± 2% 18.3ns ± 3% +79.48% (p=0.000 n=8+8) MapAccessMiss/Int64/18-16 10.3ns ± 1% 7.2ns ± 1% -29.85% (p=0.000 n=9+8) MapAccessMiss/Int64/24-16 13.9ns ± 2% 10.8ns ± 2% -22.43% (p=0.000 n=7+8) MapAccessMiss/Int64/30-16 10.1ns ± 3% 7.1ns ± 3% -29.69% (p=0.000 n=8+8) MapAccessMiss/Int64/64-16 10.3ns ± 1% 7.9ns ±12% -23.29% (p=0.000 n=7+10) MapAccessMiss/Int64/128-16 10.2ns ± 2% 7.6ns ± 8% -25.64% (p=0.000 n=10+9) MapAccessMiss/Int64/256-16 10.4ns ± 2% 7.9ns ± 7% -23.40% (p=0.000 n=9+9) MapAccessMiss/Int64/512-16 10.3ns ± 3% 7.9ns ± 8% -23.37% (p=0.000 n=9+10) MapAccessMiss/Int64/1024-16 10.4ns ± 2% 8.0ns ± 4% -22.40% (p=0.000 n=9+10) MapAccessMiss/Int64/2048-16 10.4ns ± 2% 7.9ns ± 3% -23.97% (p=0.000 n=10+10) MapAccessMiss/Int64/4096-16 10.4ns ± 2% 8.2ns ± 3% -21.67% (p=0.000 n=10+10) MapAccessMiss/Int64/8192-16 10.3ns ± 2% 8.6ns ± 3% -16.95% (p=0.000 n=10+10) MapAccessMiss/Int64/65536-16 13.5ns ± 2% 10.9ns ± 3% -19.23% (p=0.000 n=9+8) MapAccessMiss/Int32/6-16 6.17ns ± 2% 7.45ns ± 2% +20.86% (p=0.000 n=10+10) MapAccessMiss/Int32/12-16 8.50ns ± 1% 15.94ns ± 3% +87.44% (p=0.000 n=8+8) MapAccessMiss/Int32/18-16 8.49ns ± 2% 8.21ns ±30% ~ (p=0.156 n=9+10) MapAccessMiss/Int32/24-16 11.9ns ± 2% 10.8ns ± 2% -9.03% (p=0.000 n=8+7) MapAccessMiss/Int32/30-16 8.48ns ± 1% 7.53ns ±15% ~ (p=0.138 n=10+10) MapAccessMiss/Int32/64-16 8.49ns ± 1% 8.11ns ±17% -4.50% (p=0.034 n=8+10) MapAccessMiss/Int32/128-16 8.89ns ± 2% 7.86ns ±10% -11.57% (p=0.000 n=7+10) MapAccessMiss/Int32/256-16 8.72ns ± 3% 7.83ns ± 3% -10.23% (p=0.000 n=10+9) MapAccessMiss/Int32/512-16 8.79ns ± 4% 7.70ns ± 4% -12.40% (p=0.000 n=10+9) MapAccessMiss/Int32/1024-16 8.81ns ± 1% 7.74ns ± 5% -12.10% (p=0.000 n=9+10) MapAccessMiss/Int32/2048-16 9.16ns ± 1% 7.91ns ± 2% -13.64% (p=0.000 n=7+9) MapAccessMiss/Int32/4096-16 9.40ns ± 2% 8.15ns ± 3% -13.29% (p=0.000 n=9+9) MapAccessMiss/Int32/8192-16 9.47ns ± 2% 8.29ns ± 2% -12.45% (p=0.000 n=9+9) MapAccessMiss/Int32/65536-16 12.9ns ± 2% 10.6ns ± 2% -18.32% (p=0.000 n=8+10) MapAccessMiss/Str/6-16 5.69ns ± 1% 6.89ns ± 1% +21.05% (p=0.000 n=10+9) MapAccessMiss/Str/12-16 11.7ns ± 5% 10.6ns ±11% -9.74% (p=0.001 n=10+10) MapAccessMiss/Str/18-16 10.2ns ± 1% 9.6ns ± 1% -5.81% (p=0.000 n=8+8) MapAccessMiss/Str/24-16 12.5ns ±11% 11.4ns ± 9% -8.83% (p=0.011 n=10+10) MapAccessMiss/Str/30-16 10.8ns ± 4% 8.9ns ± 1% -17.66% (p=0.000 n=10+8) MapAccessMiss/Str/64-16 10.8ns ± 4% 9.3ns ± 8% -13.63% (p=0.000 n=10+10) MapAccessMiss/Str/128-16 12.0ns ± 7% 9.9ns ±10% -17.57% (p=0.000 n=10+10) MapAccessMiss/Str/256-16 11.2ns ± 3% 10.0ns ± 5% -11.09% (p=0.000 n=10+10) MapAccessMiss/Str/512-16 11.1ns ± 3% 9.6ns ± 6% -13.51% (p=0.000 n=10+10) MapAccessMiss/Str/1024-16 12.3ns ± 2% 9.7ns ± 3% -20.85% (p=0.000 n=10+10) MapAccessMiss/Str/2048-16 15.7ns ± 2% 10.1ns ± 4% -35.32% (p=0.000 n=10+10) MapAccessMiss/Str/4096-16 14.0ns ± 3% 10.2ns ± 3% -27.20% (p=0.000 n=10+10) MapAccessMiss/Str/8192-16 14.1ns ± 2% 10.9ns ± 4% -22.70% (p=0.000 n=10+10) MapAccessMiss/Str/65536-16 19.2ns ± 2% 13.6ns ± 4% -29.47% (p=0.000 n=10+10) MapAssignGrow/Int64/6-16 69.1ns ± 1% 69.5ns ± 1% ~ (p=0.133 n=9+10) MapAssignGrow/Int64/12-16 759ns ± 0% 539ns ± 0% -28.91% (p=0.000 n=7+10) MapAssignGrow/Int64/18-16 1.67µs ± 0% 1.19µs ± 1% -28.56% (p=0.000 n=10+10) MapAssignGrow/Int64/24-16 2.02µs ± 0% 1.36µs ± 0% -32.89% (p=0.000 n=10+10) MapAssignGrow/Int64/30-16 3.53µs ± 0% 2.53µs ± 0% -28.19% (p=0.000 n=8+10) MapAssignGrow/Int64/64-16 7.84µs ± 1% 5.38µs ± 1% -31.36% (p=0.000 n=10+10) MapAssignGrow/Int64/128-16 15.5µs ± 0% 10.9µs ± 0% -29.52% (p=0.000 n=9+10) MapAssignGrow/Int64/256-16 30.2µs ± 0% 21.9µs ± 0% -27.59% (p=0.000 n=10+10) MapAssignGrow/Int64/512-16 58.8µs ± 0% 43.7µs ± 0% -25.58% (p=0.000 n=10+9) MapAssignGrow/Int64/1024-16 116µs ± 0% 88µs ± 0% -24.24% (p=0.000 n=10+9) MapAssignGrow/Int64/2048-16 231µs ± 0% 175µs ± 0% -24.04% (p=0.000 n=10+10) MapAssignGrow/Int64/4096-16 458µs ± 0% 351µs ± 0% -23.49% (p=0.000 n=10+10) MapAssignGrow/Int64/8192-16 919µs ± 0% 712µs ± 0% -22.51% (p=0.000 n=10+10) MapAssignGrow/Int64/65536-16 8.69ms ± 1% 6.29ms ± 0% -27.66% (p=0.000 n=10+7) MapAssignGrow/Int32/6-16 71.7ns ± 1% 67.6ns ± 1% -5.72% (p=0.000 n=9+10) MapAssignGrow/Int32/12-16 734ns ± 0% 521ns ± 0% -29.00% (p=0.000 n=9+10) MapAssignGrow/Int32/18-16 1.61µs ± 1% 1.15µs ± 0% -28.59% (p=0.000 n=10+7) MapAssignGrow/Int32/24-16 1.97µs ± 0% 1.32µs ± 0% -33.10% (p=0.000 n=10+9) MapAssignGrow/Int32/30-16 3.41µs ± 0% 2.43µs ± 1% -28.67% (p=0.000 n=10+10) MapAssignGrow/Int32/64-16 7.45µs ± 0% 5.13µs ± 0% -31.09% (p=0.000 n=10+10) MapAssignGrow/Int32/128-16 15.0µs ± 0% 10.7µs ± 0% -28.89% (p=0.000 n=10+9) MapAssignGrow/Int32/256-16 29.8µs ± 0% 22.0µs ± 1% -26.05% (p=0.000 n=9+10) MapAssignGrow/Int32/512-16 58.1µs ± 0% 42.9µs ± 0% -26.07% (p=0.000 n=10+8) MapAssignGrow/Int32/1024-16 115µs ± 0% 86µs ± 0% -24.96% (p=0.000 n=10+10) MapAssignGrow/Int32/2048-16 226µs ± 1% 171µs ± 1% -24.44% (p=0.000 n=10+10) MapAssignGrow/Int32/4096-16 445µs ± 1% 341µs ± 0% -23.31% (p=0.000 n=10+10) MapAssignGrow/Int32/8192-16 900µs ± 0% 686µs ± 0% -23.72% (p=0.000 n=10+10) MapAssignGrow/Int32/65536-16 7.96ms ± 0% 6.11ms ± 1% -23.18% (p=0.000 n=7+10) MapAssignGrow/Str/6-16 86.5ns ± 2% 91.6ns ± 2% +5.88% (p=0.000 n=10+10) MapAssignGrow/Str/12-16 908ns ± 0% 719ns ± 0% -20.87% (p=0.000 n=10+10) MapAssignGrow/Str/18-16 2.01µs ± 0% 1.59µs ± 0% -20.98% (p=0.000 n=7+8) MapAssignGrow/Str/24-16 2.36µs ± 0% 1.79µs ± 1% -23.93% (p=0.000 n=10+10) MapAssignGrow/Str/30-16 4.17µs ± 0% 3.22µs ± 0% -22.72% (p=0.000 n=10+10) MapAssignGrow/Str/64-16 9.21µs ± 0% 6.73µs ± 0% -26.96% (p=0.000 n=7+9) MapAssignGrow/Str/128-16 18.6µs ± 0% 13.4µs ± 1% -27.60% (p=0.000 n=8+10) MapAssignGrow/Str/256-16 35.5µs ± 0% 26.8µs ± 0% -24.53% (p=0.000 n=10+10) MapAssignGrow/Str/512-16 70.6µs ± 0% 53.3µs ± 0% -24.58% (p=0.000 n=10+10) MapAssignGrow/Str/1024-16 142µs ± 1% 109µs ± 0% -23.46% (p=0.000 n=9+10) MapAssignGrow/Str/2048-16 289µs ± 1% 221µs ± 0% -23.54% (p=0.000 n=10+9) MapAssignGrow/Str/4096-16 588µs ± 0% 448µs ± 0% -23.83% (p=0.000 n=9+9) MapAssignGrow/Str/8192-16 1.27ms ± 1% 0.91ms ± 0% -28.82% (p=0.000 n=10+10) MapAssignGrow/Str/65536-16 12.4ms ± 1% 9.9ms ± 1% -20.47% (p=0.000 n=10+9) MapAssignPreAllocate/Pointer/6-16 310ns ± 1% 281ns ± 0% -9.41% (p=0.000 n=10+7) MapAssignPreAllocate/Pointer/12-16 864ns ± 0% 668ns ± 1% -22.73% (p=0.000 n=10+10) MapAssignPreAllocate/Pointer/18-16 1.31µs ± 0% 0.97µs ± 0% -25.94% (p=0.000 n=10+10) MapAssignPreAllocate/Pointer/24-16 1.77µs ± 0% 1.26µs ± 0% -29.07% (p=0.000 n=10+9) MapAssignPreAllocate/Pointer/30-16 2.16µs ± 0% 1.60µs ± 1% -26.13% (p=0.000 n=10+9) MapAssignPreAllocate/Pointer/64-16 4.73µs ± 1% 3.23µs ± 0% -31.72% (p=0.000 n=10+10) MapAssignPreAllocate/Pointer/128-16 9.18µs ± 0% 6.36µs ± 1% -30.65% (p=0.000 n=10+10) MapAssignPreAllocate/Pointer/256-16 18.0µs ± 0% 12.4µs ± 0% -30.98% (p=0.000 n=8+10) MapAssignPreAllocate/Pointer/512-16 36.2µs ± 1% 24.6µs ± 0% -32.11% (p=0.000 n=10+8) MapAssignPreAllocate/Pointer/1024-16 72.4µs ± 1% 50.3µs ± 0% -30.49% (p=0.000 n=10+10) MapAssignPreAllocate/Pointer/2048-16 145µs ± 1% 102µs ± 0% -29.93% (p=0.000 n=10+10) MapAssignPreAllocate/Pointer/4096-16 291µs ± 0% 209µs ± 1% -28.18% (p=0.000 n=10+9) MapAssignPreAllocate/Pointer/8192-16 590µs ± 0% 444µs ± 1% -24.76% (p=0.000 n=10+10) MapAssignPreAllocate/Pointer/65536-16 6.21ms ± 1% 5.68ms ± 1% -8.48% (p=0.000 n=9+9) MapAssignPreAllocate/Int64/6-16 74.7ns ± 2% 71.2ns ± 2% -4.68% (p=0.000 n=10+10) MapAssignPreAllocate/Int64/12-16 540ns ± 0% 360ns ± 0% -33.39% (p=0.000 n=9+9) MapAssignPreAllocate/Int64/18-16 811ns ± 0% 525ns ± 0% -35.20% (p=0.000 n=9+9) MapAssignPreAllocate/Int64/24-16 1.16µs ± 0% 0.68µs ± 1% -41.94% (p=0.000 n=9+10) MapAssignPreAllocate/Int64/30-16 1.33µs ± 1% 0.87µs ± 1% -34.60% (p=0.000 n=10+10) MapAssignPreAllocate/Int64/64-16 2.97µs ± 1% 1.75µs ± 0% -41.01% (p=0.000 n=10+10) MapAssignPreAllocate/Int64/128-16 5.73µs ± 0% 3.39µs ± 0% -40.79% (p=0.000 n=9+8) MapAssignPreAllocate/Int64/256-16 11.0µs ± 0% 6.5µs ± 0% -40.50% (p=0.000 n=10+8) MapAssignPreAllocate/Int64/512-16 21.8µs ± 0% 12.9µs ± 0% -40.91% (p=0.000 n=10+10) MapAssignPreAllocate/Int64/1024-16 43.6µs ± 0% 26.4µs ± 0% -39.42% (p=0.000 n=10+10) MapAssignPreAllocate/Int64/2048-16 88.0µs ± 1% 52.5µs ± 0% -40.40% (p=0.000 n=10+8) MapAssignPreAllocate/Int64/4096-16 175µs ± 0% 108µs ± 0% -38.60% (p=0.000 n=9+10) MapAssignPreAllocate/Int64/8192-16 362µs ± 0% 224µs ± 1% -38.10% (p=0.000 n=8+10) MapAssignPreAllocate/Int64/65536-16 3.78ms ± 0% 3.21ms ± 2% -15.14% (p=0.000 n=10+10) MapAssignPreAllocate/Int32/6-16 72.8ns ± 1% 69.2ns ± 1% -4.98% (p=0.000 n=9+10) MapAssignPreAllocate/Int32/12-16 503ns ± 0% 342ns ± 0% -32.02% (p=0.000 n=10+9) MapAssignPreAllocate/Int32/18-16 770ns ± 1% 498ns ± 1% -35.29% (p=0.000 n=10+8) MapAssignPreAllocate/Int32/24-16 1.11µs ± 1% 0.65µs ± 0% -41.79% (p=0.000 n=10+8) MapAssignPreAllocate/Int32/30-16 1.26µs ± 0% 0.81µs ± 1% -35.71% (p=0.000 n=10+8) MapAssignPreAllocate/Int32/64-16 2.79µs ± 1% 1.63µs ± 0% -41.60% (p=0.000 n=10+9) MapAssignPreAllocate/Int32/128-16 5.60µs ± 1% 3.39µs ± 1% -39.52% (p=0.000 n=10+10) MapAssignPreAllocate/Int32/256-16 11.1µs ± 1% 6.7µs ± 1% -39.66% (p=0.000 n=10+10) MapAssignPreAllocate/Int32/512-16 21.6µs ± 0% 12.2µs ± 0% -43.30% (p=0.000 n=10+10) MapAssignPreAllocate/Int32/1024-16 42.4µs ± 1% 24.2µs ± 0% -42.91% (p=0.000 n=10+10) MapAssignPreAllocate/Int32/2048-16 84.9µs ± 0% 50.0µs ± 0% -41.08% (p=0.000 n=10+8) MapAssignPreAllocate/Int32/4096-16 173µs ± 0% 104µs ± 0% -40.04% (p=0.000 n=10+10) MapAssignPreAllocate/Int32/8192-16 347µs ± 1% 212µs ± 1% -39.07% (p=0.000 n=10+8) MapAssignPreAllocate/Int32/65536-16 3.69ms ± 1% 3.07ms ± 2% -16.96% (p=0.000 n=10+10) MapAssignPreAllocate/Str/6-16 89.6ns ± 2% 94.6ns ± 1% +5.49% (p=0.000 n=9+7) MapAssignPreAllocate/Str/12-16 641ns ± 1% 504ns ± 0% -21.37% (p=0.000 n=9+9) MapAssignPreAllocate/Str/18-16 1.01µs ± 0% 0.76µs ± 0% -24.98% (p=0.000 n=7+8) MapAssignPreAllocate/Str/24-16 1.36µs ± 0% 0.96µs ± 0% -29.74% (p=0.000 n=9+9) MapAssignPreAllocate/Str/30-16 1.68µs ± 0% 1.22µs ± 0% -27.15% (p=0.000 n=10+9) MapAssignPreAllocate/Str/64-16 3.82µs ± 0% 2.43µs ± 0% -36.29% (p=0.000 n=10+10) MapAssignPreAllocate/Str/128-16 7.60µs ± 0% 4.66µs ± 0% -38.67% (p=0.000 n=9+10) MapAssignPreAllocate/Str/256-16 13.8µs ± 0% 9.2µs ± 0% -33.55% (p=0.000 n=10+10) MapAssignPreAllocate/Str/512-16 27.5µs ± 0% 18.2µs ± 0% -33.89% (p=0.000 n=10+9) MapAssignPreAllocate/Str/1024-16 55.4µs ± 0% 37.8µs ± 0% -31.90% (p=0.000 n=10+10) MapAssignPreAllocate/Str/2048-16 111µs ± 0% 77µs ± 0% -30.20% (p=0.000 n=9+9) MapAssignPreAllocate/Str/4096-16 226µs ± 0% 159µs ± 0% -29.57% (p=0.000 n=10+9) MapAssignPreAllocate/Str/8192-16 481µs ± 0% 368µs ± 1% -23.34% (p=0.000 n=9+10) MapAssignPreAllocate/Str/65536-16 4.83ms ± 1% 4.11ms ± 1% -14.86% (p=0.000 n=10+10) MapAssignReuse/Pointer/6-16 330ns ± 2% 290ns ± 1% -12.18% (p=0.000 n=10+10) MapAssignReuse/Pointer/12-16 757ns ± 1% 550ns ± 1% -27.35% (p=0.000 n=9+9) MapAssignReuse/Pointer/18-16 1.16µs ± 1% 0.81µs ± 1% -30.13% (p=0.000 n=9+10) MapAssignReuse/Pointer/24-16 1.61µs ± 1% 1.08µs ± 1% -32.59% (p=0.000 n=8+10) MapAssignReuse/Pointer/30-16 1.91µs ± 2% 1.31µs ± 1% -31.72% (p=0.000 n=10+9) MapAssignReuse/Pointer/64-16 4.12µs ± 1% 2.75µs ± 1% -33.26% (p=0.000 n=8+9) MapAssignReuse/Pointer/128-16 8.22µs ± 2% 5.50µs ± 1% -33.03% (p=0.000 n=10+10) MapAssignReuse/Pointer/256-16 16.4µs ± 1% 11.0µs ± 1% -33.01% (p=0.000 n=9+10) MapAssignReuse/Pointer/512-16 33.1µs ± 2% 21.9µs ± 1% -33.72% (p=0.000 n=10+10) MapAssignReuse/Pointer/1024-16 66.7µs ± 2% 44.8µs ± 1% -32.78% (p=0.000 n=10+9) MapAssignReuse/Pointer/2048-16 132µs ± 2% 92µs ± 1% -30.11% (p=0.000 n=10+10) MapAssignReuse/Pointer/4096-16 268µs ± 1% 187µs ± 1% -30.25% (p=0.000 n=9+10) MapAssignReuse/Pointer/8192-16 546µs ± 4% 389µs ± 1% -28.67% (p=0.000 n=10+10) MapAssignReuse/Pointer/65536-16 5.14ms ± 3% 4.18ms ± 1% -18.60% (p=0.000 n=10+10) MapAssignReuse/Int64/6-16 81.1ns ± 1% 74.8ns ± 2% -7.75% (p=0.000 n=10+10) MapAssignReuse/Int64/12-16 413ns ± 3% 152ns ± 1% -63.04% (p=0.000 n=9+10) MapAssignReuse/Int64/18-16 429ns ± 3% 223ns ± 1% -48.00% (p=0.000 n=8+10) MapAssignReuse/Int64/24-16 1.01µs ± 4% 0.30µs ± 1% -69.93% (p=0.000 n=10+9) MapAssignReuse/Int64/30-16 618ns ± 3% 358ns ± 2% -42.00% (p=0.000 n=9+10) MapAssignReuse/Int64/64-16 1.31µs ± 1% 0.75µs ± 2% -42.74% (p=0.000 n=8+8) MapAssignReuse/Int64/128-16 2.64µs ± 3% 1.51µs ± 2% -42.96% (p=0.000 n=9+10) MapAssignReuse/Int64/256-16 5.23µs ± 4% 3.00µs ± 2% -42.59% (p=0.000 n=10+10) MapAssignReuse/Int64/512-16 10.3µs ± 3% 6.0µs ± 1% -41.70% (p=0.000 n=10+8) MapAssignReuse/Int64/1024-16 21.0µs ± 2% 12.1µs ± 2% -42.09% (p=0.000 n=10+10) MapAssignReuse/Int64/2048-16 42.2µs ± 2% 25.5µs ± 2% -39.56% (p=0.000 n=10+10) MapAssignReuse/Int64/4096-16 85.3µs ± 3% 51.3µs ± 2% -39.88% (p=0.000 n=10+10) MapAssignReuse/Int64/8192-16 176µs ± 3% 107µs ± 2% -39.54% (p=0.000 n=10+10) MapAssignReuse/Int64/65536-16 1.73ms ± 3% 1.16ms ± 2% -33.33% (p=0.000 n=10+10) MapAssignReuse/Int32/6-16 79.6ns ± 1% 72.4ns ± 2% -9.14% (p=0.000 n=10+10) MapAssignReuse/Int32/12-16 356ns ± 7% 150ns ± 1% -57.73% (p=0.000 n=9+10) MapAssignReuse/Int32/18-16 394ns ± 5% 217ns ± 2% -44.82% (p=0.000 n=9+9) MapAssignReuse/Int32/24-16 1.01µs ± 7% 0.29µs ± 2% -70.98% (p=0.000 n=10+10) MapAssignReuse/Int32/30-16 618ns ± 5% 348ns ± 2% -43.68% (p=0.000 n=10+10) MapAssignReuse/Int32/64-16 1.29µs ± 4% 0.73µs ± 2% -43.23% (p=0.000 n=10+10) MapAssignReuse/Int32/128-16 2.61µs ± 3% 1.46µs ± 2% -43.88% (p=0.000 n=10+10) MapAssignReuse/Int32/256-16 5.21µs ± 3% 2.92µs ± 2% -43.90% (p=0.000 n=10+10) MapAssignReuse/Int32/512-16 10.2µs ± 3% 5.9µs ± 1% -42.54% (p=0.000 n=10+9) MapAssignReuse/Int32/1024-16 20.5µs ± 3% 11.8µs ± 1% -42.72% (p=0.000 n=10+9) MapAssignReuse/Int32/2048-16 41.7µs ± 3% 24.8µs ± 2% -40.50% (p=0.000 n=9+10) MapAssignReuse/Int32/4096-16 84.5µs ± 2% 50.8µs ± 1% -39.93% (p=0.000 n=10+9) MapAssignReuse/Int32/8192-16 170µs ± 1% 103µs ± 2% -39.52% (p=0.000 n=10+8) MapAssignReuse/Int32/65536-16 1.68ms ± 3% 1.12ms ± 2% -33.25% (p=0.000 n=10+10) MapAssignReuse/Str/6-16 96.8ns ± 3% 102.0ns ± 1% +5.33% (p=0.000 n=10+10) MapAssignReuse/Str/12-16 491ns ± 3% 202ns ± 1% -58.99% (p=0.000 n=8+9) MapAssignReuse/Str/18-16 500ns ± 4% 292ns ± 2% -41.68% (p=0.000 n=10+10) MapAssignReuse/Str/24-16 1.13µs ± 3% 0.40µs ± 1% -64.33% (p=0.000 n=9+10) MapAssignReuse/Str/30-16 783ns ± 5% 480ns ± 2% -38.77% (p=0.000 n=10+10) MapAssignReuse/Str/64-16 1.53µs ± 2% 1.02µs ± 2% -33.00% (p=0.000 n=10+10) MapAssignReuse/Str/128-16 3.06µs ± 1% 2.04µs ± 2% -33.22% (p=0.000 n=10+10) MapAssignReuse/Str/256-16 6.15µs ± 3% 4.08µs ± 2% -33.59% (p=0.000 n=9+10) MapAssignReuse/Str/512-16 12.3µs ± 3% 8.2µs ± 1% -33.49% (p=0.000 n=10+10) MapAssignReuse/Str/1024-16 25.3µs ± 2% 17.1µs ± 2% -32.55% (p=0.000 n=10+10) MapAssignReuse/Str/2048-16 50.7µs ± 3% 34.6µs ± 2% -31.79% (p=0.000 n=10+10) MapAssignReuse/Str/4096-16 103µs ± 2% 70µs ± 2% -31.98% (p=0.000 n=10+10) MapAssignReuse/Str/8192-16 218µs ± 1% 157µs ± 2% -28.32% (p=0.000 n=9+10) MapAssignReuse/Str/65536-16 2.03ms ± 1% 1.64ms ± 2% -19.62% (p=0.000 n=10+9) Benchmark-2The benchmarks from runtime. INFOname old time/op new time/op delta MegMap-16 9.01ns ± 1% 8.83ns ± 1% -2.00% (p=0.000 n=8+8) MegOneMap-16 4.59ns ± 1% 9.21ns ± 2% +100.52% (p=0.000 n=9+10) MegEqMap-16 20.4µs ± 2% 20.1µs ± 2% -1.49% (p=0.011 n=9+9) MegEmptyMap-16 2.18ns ± 3% 2.06ns ± 1% -5.42% (p=0.000 n=10+10) SmallStrMap-16 8.27ns ± 1% 7.17ns ± 4% -13.31% (p=0.000 n=8+10) MapStringKeysEight_16-16 6.89ns ± 3% 7.79ns ± 1% +13.09% (p=0.000 n=9+8) MapStringKeysEight_32-16 7.75ns ± 5% 7.97ns ± 0% +2.83% (p=0.000 n=10+7) MapStringKeysEight_64-16 7.66ns ± 3% 8.76ns ± 1% +14.43% (p=0.000 n=10+7) MapStringKeysEight_1M-16 7.72ns ± 3% 16483.00ns ± 1% +213363.36% (p=0.000 n=10+8) IntMap-16 5.21ns ± 1% 6.42ns ±16% +23.30% (p=0.000 n=8+10) MapFirst/1-16 2.68ns ± 1% 2.45ns ± 1% -8.78% (p=0.000 n=10+9) MapFirst/2-16 2.68ns ± 1% 2.43ns ± 1% -9.34% (p=0.000 n=9+9) MapFirst/3-16 2.66ns ± 2% 2.44ns ± 1% -7.98% (p=0.000 n=10+9) MapFirst/4-16 2.68ns ± 1% 2.44ns ± 0% -8.67% (p=0.000 n=8+9) MapFirst/5-16 2.69ns ± 1% 2.45ns ± 0% -8.73% (p=0.000 n=8+8) MapFirst/6-16 2.68ns ± 0% 2.44ns ± 1% -8.78% (p=0.000 n=8+9) MapFirst/7-16 2.69ns ± 1% 2.45ns ± 1% -9.10% (p=0.000 n=10+9) MapFirst/8-16 2.67ns ± 1% 4.79ns ± 1% +79.47% (p=0.000 n=10+9) MapFirst/9-16 4.66ns ± 1% 4.75ns ± 2% +1.98% (p=0.000 n=10+9) MapFirst/10-16 4.68ns ± 2% 4.73ns ± 1% +1.19% (p=0.016 n=9+10) MapFirst/11-16 4.69ns ± 2% 4.78ns ± 1% +1.93% (p=0.000 n=9+8) MapFirst/12-16 4.64ns ± 2% 4.75ns ± 2% +2.47% (p=0.005 n=10+10) MapFirst/13-16 4.69ns ± 1% 4.78ns ± 1% +1.86% (p=0.000 n=8+9) MapFirst/14-16 4.68ns ± 2% 4.79ns ± 1% +2.34% (p=0.000 n=10+8) MapFirst/15-16 4.67ns ± 1% 6.90ns ± 2% +47.88% (p=0.000 n=8+10) MapFirst/16-16 4.65ns ± 2% 6.95ns ± 1% +49.60% (p=0.000 n=10+9) MapMid/1-16 2.68ns ± 1% 2.62ns ± 1% -2.15% (p=0.000 n=10+9) MapMid/2-16 3.23ns ± 1% 2.95ns ± 1% -8.59% (p=0.000 n=10+9) MapMid/3-16 3.21ns ± 1% 2.94ns ± 0% -8.33% (p=0.000 n=10+7) MapMid/4-16 3.64ns ± 2% 3.46ns ± 1% -4.84% (p=0.000 n=10+8) MapMid/5-16 3.67ns ± 1% 3.47ns ± 1% -5.46% (p=0.000 n=10+8) MapMid/6-16 4.15ns ± 2% 3.93ns ± 1% -5.25% (p=0.000 n=10+9) MapMid/7-16 4.19ns ± 1% 3.91ns ± 1% -6.55% (p=0.000 n=10+10) MapMid/8-16 4.84ns ± 2% 5.76ns ±11% +19.02% (p=0.000 n=10+10) MapMid/9-16 5.66ns ±10% 5.83ns ± 9% ~ (p=0.393 n=10+10) MapMid/10-16 5.81ns ± 5% 6.20ns ± 4% +6.67% (p=0.001 n=8+7) MapMid/11-16 5.80ns ± 4% 6.17ns ±11% ~ (p=0.278 n=9+10) MapMid/12-16 5.87ns ±10% 6.26ns ± 2% +6.48% (p=0.008 n=9+7) MapMid/13-16 6.15ns ±13% 5.92ns ± 8% ~ (p=0.315 n=10+8) MapMid/14-16 5.29ns ± 8% 6.18ns ±14% +16.89% (p=0.001 n=9+10) MapMid/15-16 5.44ns ±12% 7.04ns ± 2% +29.37% (p=0.000 n=10+10) MapMid/16-16 5.91ns ±11% 7.18ns ± 3% +21.44% (p=0.000 n=10+9) MapLast/1-16 2.69ns ± 1% 2.64ns ± 2% -2.00% (p=0.000 n=8+10) MapLast/2-16 3.22ns ± 2% 2.93ns ± 1% -9.05% (p=0.000 n=9+10) MapLast/3-16 3.66ns ± 2% 3.47ns ± 1% -5.17% (p=0.000 n=10+8) MapLast/4-16 4.20ns ± 1% 3.88ns ± 2% -7.54% (p=0.000 n=8+9) MapLast/5-16 4.85ns ± 2% 4.39ns ± 2% -9.53% (p=0.000 n=9+9) MapLast/6-16 5.33ns ± 1% 5.10ns ± 8% -4.27% (p=0.017 n=9+10) MapLast/7-16 5.79ns ± 1% 5.33ns ± 2% -7.93% (p=0.000 n=9+8) MapLast/8-16 6.28ns ± 1% 6.55ns ±17% ~ (p=0.704 n=9+10) MapLast/9-16 6.90ns ±18% 6.87ns ±12% ~ (p=1.000 n=9+9) MapLast/10-16 6.83ns ±15% 7.32ns ± 7% +7.10% (p=0.010 n=9+8) MapLast/11-16 7.84ns ± 3% 7.46ns ± 5% -4.78% (p=0.006 n=7+8) MapLast/12-16 7.58ns ± 9% 7.58ns ±12% ~ (p=0.743 n=8+9) MapLast/13-16 9.15ns ±25% 14.06ns ± 2% +53.70% (p=0.000 n=10+8) MapLast/14-16 6.31ns ±15% 7.95ns ± 8% +25.94% (p=0.000 n=9+8) MapLast/15-16 6.20ns ±16% 7.11ns ± 2% +14.59% (p=0.004 n=10+10) MapLast/16-16 6.78ns ±19% 7.19ns ± 1% ~ (p=0.139 n=9+8) MapCycle-16 11.5ns ± 2% 15.8ns ± 2% +36.83% (p=0.000 n=10+10) RepeatedLookupStrMapKey32-16 9.37ns ± 2% 7.81ns ± 1% -16.67% (p=0.000 n=9+9) RepeatedLookupStrMapKey1M-16 16.6µs ± 1% 16.6µs ± 2% ~ (p=0.604 n=10+9) MakeMap/[Byte]Byte-16 120ns ± 1% 116ns ± 1% -3.61% (p=0.000 n=10+10) MakeMap/[Int]Int-16 164ns ± 1% 160ns ± 0% -2.36% (p=0.000 n=10+9) NewEmptyMap-16 4.32ns ± 1% 4.55ns ± 3% +5.26% (p=0.000 n=10+10) NewSmallMap-16 19.4ns ± 2% 24.1ns ± 1% +24.31% (p=0.000 n=10+10) MapIter-16 73.7ns ± 2% 82.3ns ± 3% +11.60% (p=0.000 n=10+10) MapIterEmpty-16 3.65ns ± 1% 4.11ns ± 2% +12.80% (p=0.000 n=9+9) SameLengthMap-16 3.14ns ± 2% 3.38ns ± 2% +7.50% (p=0.000 n=9+10) BigKeyMap-16 10.0ns ± 4% 11.9ns ± 1% +18.84% (p=0.000 n=10+10) BigValMap-16 10.0ns ± 3% 11.9ns ± 2% +18.74% (p=0.000 n=10+10) SmallKeyMap-16 8.51ns ± 1% 9.91ns ± 1% +16.44% (p=0.000 n=10+9) MapPopulate/1-16 10.4ns ± 1% 14.4ns ± 2% +38.28% (p=0.000 n=9+9) MapPopulate/10-16 608ns ± 1% 495ns ± 1% -18.58% (p=0.000 n=10+10) MapPopulate/100-16 9.32µs ± 1% 6.28µs ± 1% -32.59% (p=0.000 n=9+10) MapPopulate/1000-16 111µs ± 0% 86µs ± 0% -22.28% (p=0.000 n=10+10) MapPopulate/10000-16 965µs ± 0% 745µs ± 1% -22.78% (p=0.000 n=10+10) MapPopulate/100000-16 10.4ms ± 1% 7.4ms ± 0% -28.46% (p=0.000 n=10+10) ComplexAlgMap-16 20.6ns ± 1% 22.9ns ± 1% +10.86% (p=0.000 n=9+10) GoMapClear/Reflexive/1-16 19.3ns ± 1% 17.9ns ± 1% -7.39% (p=0.000 n=10+10) GoMapClear/Reflexive/10-16 21.2ns ± 1% 18.9ns ± 2% -10.91% (p=0.000 n=10+9) GoMapClear/Reflexive/100-16 37.3ns ± 1% 40.2ns ± 1% +7.90% (p=0.000 n=9+9) GoMapClear/Reflexive/1000-16 345ns ± 2% 433ns ± 1% +25.73% (p=0.000 n=10+10) GoMapClear/Reflexive/10000-16 2.62µs ± 1% 3.94µs ± 1% +50.33% (p=0.000 n=10+10) GoMapClear/NonReflexive/1-16 80.2ns ± 2% 62.8ns ± 2% -21.61% (p=0.000 n=9+9) GoMapClear/NonReflexive/10-16 92.2ns ± 2% 72.5ns ± 2% -21.42% (p=0.000 n=10+10) GoMapClear/NonReflexive/100-16 216ns ± 2% 101ns ± 1% -53.49% (p=0.000 n=9+8) GoMapClear/NonReflexive/1000-16 2.13µs ± 1% 0.34µs ± 2% -84.02% (p=0.000 n=10+9) GoMapClear/NonReflexive/10000-16 16.4µs ± 1% 2.1µs ± 1% -86.96% (p=0.000 n=10+8) MapStringConversion/32/simple-16 7.79ns ± 1% 12.22ns ± 1% +56.92% (p=0.000 n=9+9) MapStringConversion/32/struct-16 7.72ns ± 1% 12.32ns ± 2% +59.70% (p=0.000 n=9+10) MapStringConversion/32/array-16 7.72ns ± 3% 12.00ns ± 3% +55.34% (p=0.000 n=10+9) MapStringConversion/64/simple-16 7.47ns ± 1% 11.77ns ± 3% +57.65% (p=0.000 n=9+10) MapStringConversion/64/struct-16 7.49ns ± 1% 11.94ns ± 2% +59.38% (p=0.000 n=9+10) MapStringConversion/64/array-16 7.50ns ± 1% 11.92ns ± 2% +58.89% (p=0.000 n=9+10) MapInterfaceString-16 11.2ns ± 5% 13.7ns ±61% ~ (p=1.000 n=8+10) MapInterfacePtr-16 10.8ns ±29% 11.3ns ±52% ~ (p=0.684 n=10+10) NewEmptyMapHintLessThan8-16 6.54ns ± 1% 6.04ns ± 1% -7.56% (p=0.000 n=9+9) NewEmptyMapHintGreaterThan8-16 271ns ± 2% 268ns ± 1% ~ (p=0.062 n=10+9) MapPop100-16 10.8µs ± 6% 7.2µs ±14% -33.41% (p=0.000 n=10+10) MapPop1000-16 170µs ± 1% 101µs ± 1% -40.54% (p=0.000 n=9+9) MapPop10000-16 3.35ms ± 3% 1.74ms ± 4% -48.15% (p=0.000 n=10+9) MapAssign/Int32/256-16 10.1ns ± 4% 8.9ns ± 1% -12.13% (p=0.000 n=10+9) MapAssign/Int32/65536-16 22.0ns ± 1% 12.9ns ± 1% -41.11% (p=0.000 n=10+8) MapAssign/Int64/256-16 10.7ns ± 2% 8.6ns ± 2% -19.79% (p=0.000 n=8+9) MapAssign/Int64/65536-16 23.2ns ± 2% 13.1ns ± 2% -43.39% (p=0.000 n=10+9) MapAssign/Str/256-16 16.6ns ± 4% 13.1ns ± 3% -21.18% (p=0.000 n=10+9) MapAssign/Str/65536-16 27.5ns ± 3% 20.5ns ± 3% -25.66% (p=0.000 n=10+9) MapOperatorAssign/Int32/256-16 9.89ns ± 2% 8.69ns ± 3% -12.15% (p=0.000 n=9+9) MapOperatorAssign/Int32/65536-16 21.9ns ± 1% 12.8ns ± 3% -41.52% (p=0.000 n=10+10) MapOperatorAssign/Int64/256-16 10.9ns ± 3% 8.6ns ± 1% -21.20% (p=0.000 n=9+8) MapOperatorAssign/Int64/65536-16 22.4ns ± 0% 13.2ns ± 2% -41.15% (p=0.000 n=7+9) MapOperatorAssign/Str/256-16 1.43µs ± 2% 1.40µs ± 1% -1.93% (p=0.000 n=10+10) MapOperatorAssign/Str/65536-16 234ns ± 3% 237ns ± 4% ~ (p=0.138 n=10+10) MapAppendAssign/Int32/256-16 21.0ns ± 2% 20.1ns ± 4% -4.43% (p=0.000 n=8+10) MapAppendAssign/Int32/65536-16 38.5ns ± 2% 33.9ns ± 2% -12.00% (p=0.000 n=10+10) MapAppendAssign/Int64/256-16 21.4ns ± 3% 20.4ns ± 6% -4.26% (p=0.008 n=9+10) MapAppendAssign/Int64/65536-16 39.8ns ± 2% 34.9ns ± 1% -12.40% (p=0.000 n=10+8) MapAppendAssign/Str/256-16 52.8ns ± 7% 52.1ns ± 5% ~ (p=0.739 n=10+10) MapAppendAssign/Str/65536-16 66.5ns ± 2% 67.7ns ± 1% +1.94% (p=0.004 n=9+9) MapDelete/Int32/100-16 27.7ns ± 2% 17.8ns ± 2% -35.77% (p=0.000 n=10+10) MapDelete/Int32/1000-16 22.6ns ± 2% 12.8ns ± 1% -43.31% (p=0.000 n=9+9) MapDelete/Int32/10000-16 25.0ns ± 2% 15.2ns ± 1% -39.33% (p=0.000 n=7+10) MapDelete/Int64/100-16 29.8ns ± 3% 17.6ns ± 2% -40.98% (p=0.000 n=10+9) MapDelete/Int64/1000-16 23.3ns ± 2% 13.1ns ± 1% -43.77% (p=0.000 n=10+8) MapDelete/Int64/10000-16 25.6ns ± 2% 16.1ns ± 1% -37.13% (p=0.000 n=9+9) MapDelete/Str/100-16 38.4ns ± 4% 24.8ns ± 2% -35.29% (p=0.000 n=10+10) MapDelete/Str/1000-16 30.8ns ± 2% 20.7ns ± 1% -32.72% (p=0.000 n=10+10) MapDelete/Str/10000-16 35.8ns ± 2% 24.4ns ± 2% -31.95% (p=0.000 n=10+10) MapDelete/Pointer/100-16 30.7ns ± 2% 21.0ns ± 1% -31.69% (p=0.000 n=10+10) MapDelete/Pointer/1000-16 24.5ns ± 2% 16.6ns ± 1% -32.37% (p=0.000 n=10+10) MapDelete/Pointer/10000-16 27.1ns ± 3% 19.8ns ± 2% -26.87% (p=0.000 n=10+10) |
|
Change https://go.dev/cl/426614 mentions this issue: |
|
cc @golang/runtime |
seankhliao
added
the
NeedsInvestigation
|
cc @randall77 Overall this sounds great. The main concern I have up front here is that growing happens all at once rather than incrementally. Could you explain more why that is and if it is possible to do so incrementally? |
|
Do iterators get invalidated when the map is modified? This is common for open-addressing hash tables. An important feature of the current map is that you can modify the map while iterating and the iterator still works. |
I checked the implementation and it does look like iterators stay valid when the map is modified. |
|
The semantics of Go iterators basically mean you can't move items in the map data structure once you've placed them. So, for instance, you can't move a valid item to fill in a deleted one - you need to always use a tombstone. But other than that, iterators don't add much of a restriction to the implementation. |
|
This is really exciting. Thanks for doing all of this work.
There are many slowdowns in these benchmarks, some quite significant. Do you understand the source of these slowdowns? Is it something that can be fixed?
This is an important property to maintain because it can significantly impact the tail latency of services. We've done a lot of work throughout the Go runtime over the years to avoid unpredictable performance spikes and it would be a shame to regress here. I'm not personally familiar with the details of SwissTable. Is incremental rehashing something that can be done and just hasn't been implemented yet, or is it fundamentally difficult? If it's fundamentally difficult to do directly in SwissTable, I've been kickaround the the idea of adopting aspects of extendible hashing for a while, which could be a general solution to incremental resizing. I should probably just file an "idea" issue about this, but I'll try to lay it out here. The idea is that, once a hash table grows beyond some size threshold based on bounding how long it takes to resize a map, you switch to a two-level scheme. I'm guessing that threshold would be around 4MiB, but that would have to be determined experimentally. In the two-level scheme, the top level is an index array keyed by the top k bits of the hash, which then points to map shards, which are individual SwissTables. The Wikipedia article I linked has some good diagrams for visualizing this. Each shard s contains only the items whose top js <= k bits are all identical, and multiple index entries will point to the same shard if that shard's js < k. When an individual shard overflows the threshold size, you split just that shard into two new shards s1 and s2 with js1 = js2 = js+1, and update the index. If js1 > k, then you also double the index array in order to increase k by 1. This would enable incremental growth and bound the resize cost paid by any single map operation. Iterators can continue to work while a map is resized by holding on to the pre-resize shard they are currently iterating (let the GC collect this once all iterators have moved past it). It would also reduce overall memory footprint because the total memory size of a map doesn't have to be a power of two (even if each individual shard is). It would result in less memory fragmentation caused by large maps because it limits the size of each individual allocation. And it would address problems with our current map resizing algorithm where, if a map is grown to a size that puts it into resizing mode, but then modifications stop and the application switches to only reading from the map, accesses are relatively slow and the map continues to consume additional memory (#51410). |
|
I placed this in the Unplanned milestone for now since it seems like there's still a bunch to discuss before landing it (it's non-committal). I assigned it to @zhangyunhao116 since you're driving the effort at the moment. |
|
Thanks for your advice! @aclements We hope to get more feedback from the community and the Go team.
Many reasons affect the performance degradation of these benchmarks(for basic version):
I think it is fundamentally difficult since a single empty slot in SwissTable can break the search process(https://faultlore.com/blah/hashbrown-tldr/#deletion).
That sounds like a great idea! I'll think about this idea carefully later. |
|
I like to use small maps for many purposes, so it worries me a bit that SwissTable seems to perform worse for them. Maybe a special case is needed for small maps? |
Yes, but it may have to consider carefully(optimizing the small map may cause performance impact in other situations). SwissTable does not always cause worse performance for small maps, and we can see the cases in the benchmark(e.g., in most key miss scenarios, the performance improves instead). Could you please tell me the cases where the small maps are often used? Some of the slowdowns in the benchmarks can be fixed. PatchSet 10 should fix the slowdown of |
|
Well I use small maps often in stead of case statements, or for mapping a few input file names to processors, or for lookups such as os.Substitute can do. I am looking forward to the improved results. |
|
Thanks for the detailed breakdown of the performance losses! My take is that there are two things that need to happen on that front:
You mentioned that SwissTable is better for the "miss" case than the "hit" case. If I understand your layout correctly (I haven't looked closely), you're already packing the tophash and the data together, so you'd expect one cache miss for a hit. Is that right? Does the SIMD version of SwissTable level the field between the hit and miss cases at all? |
|
Stepping back, I think there are three high-level potential concerns to address with this: Incremental resizing. We've already been discussing this above. :) DoS prevention. Currently, Go maps have fairly strong denial-of-service properties. The main concern is if an attacker can control the placement of items to induce collisions, leading to n^2 behavior. Having a good, per-map seed that is factored into the hash computation (not just combined at the end) goes a long way toward preventing this. Beyond this, the concern becomes leaking too much information about long-lived maps such that an attacker could reconstruct the seed. Having non-deterministic iteration order helps a lot here. A third problem is that copying elements one-by-one from one map into another, newly allocated map can easily go n^2, but having a per-map seed also mitigates that. I think all of these mechanisms are still in place in your implementation, but wanted to confirm that. SIMD, specifically how to actually use SIMD instructions in the map implementation. The most straightforward approach is to write the SIMD bits in assembly, but that has significant maintainability problems and may even perform worse because of the cost of transitioning to and from assembly. So, that suggests adding compiler intrinsics. General SIMD intrinsics are a wide-open area right now and I don't think we want to block the map implementation on solving that problem. :) A lot of proposals have floated around (admittedly, some of which have not been published). We could start trying them out just in the runtime. Or, if I understand correctly, SwissTable doesn't need very many or any particularly fancy SIMD operations, so maybe we just hack in some intrinsics (which we could use as a starting point for more general exploration). |
Another option that would help alleviate most n^2 concerns is having a per-map seed that changes when the map grows/shrinks. |
This is challenging with incremental map growth. FWIW, I believe we do opportunistically reset the seed when the length hits zero. |
|
Thanks for all of the feedback :)
PatchSet-11 adding the optimization of There are still two optimizations that can be added
Indeed. We can add some new benchmarks to it. I feel like adding the new one into https://github.com/zhangyunhao116/gomapbench first and then migrating some of it into the runtime is a good idea.
Yes! All of these mechanisms are still in the current implementation, including the non-deterministic iteration and resetting the map's seed when the length hits zero.
We planned to write the SIMD directly in assembly, but the experiments outside the runtime show that their performance is almost the same. I wonder if we can use ABIInternal can make the assembly run faster? I think SIMD intrinsics is a great idea! Agree that we can start trying SIMD intrinsics just in the runtime. After some basic optimizations are done, I will start to implement the SIMD version :) Most SIMD in the SwissTable is not very tricky except the |
|
Another challenge of SIMD is that its default bucket size is 16(the basic version is 8, which is the same as before), which means memory consumption may increase significantly in some cases. |
Sounds good. Possibly we should fold your package into bent, so they run as part of the performance dashboard.
Perfect!
Yeah, this doesn't surprise me if you're jumping to assembly just to run a few instructions. ABIInternal would certainly improve that, since it would probably allow the entire call to be registerized. I'm not sure exactly what operations you need, but if they're taking or returning vector values there's still going to be some overhead since ABIInternal doesn't know anything about passing vector registers. I'm not sure how much effort it is to add a few one-off SIMD intrinsics. You just need SSE, right, not AVX? The operations themselves are probably not hard to add. The register allocator already knows about the XMM registers, but only as non-vector 64-bit floating-point. That mostly shouldn't matter unless it decides to spill one (thanks @cherrymui for pointing that out). If you need AVX, things may get a lot more complicated, or maybe not since the YMM registers are just aliased onto the XMM registers, so we would just call them X registers anyway.
It's hard to say how concerning this is. Sure, if an application has a huge number of small maps, perhaps with large key or value types (though they can't go over 128 bytes each), this could add up to a lot. But the relative overhead of this decreases as map size increases. I'll see if we can get some data on this. Are you considering the trick described in this video, where you need to allocate 16 bytes of control (or a little more for unaligned groups) but the actual bucket array can be smaller? The net result of that might actually be that the tables would be smaller than current tables. |
This enables manual inlining Map.Get/table.getWithoutKey to create a simple fast path with no calls. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: Ic208dd4c02c7554f312b85b5fadccaf82b23545c Reviewed-on: https://go-review.googlesource.com/c/go/+/616455 LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Keith Randall <khr@google.com> Auto-Submit: Michael Pratt <mpratt@google.com>
When given a hint size, set the initial capacity large enough to avoid requiring growth in the average case. When not given a hint (or given 0), don't allocate anything at all. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: I8844fc652b8d2d4e5136cd56f7e78999a07fe381 Reviewed-on: https://go-review.googlesource.com/c/go/+/616457 LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Keith Randall <khr@google.com> Auto-Submit: Michael Pratt <mpratt@google.com>
Rather than importing runtime directly, linkname the functions from runtime. This allows importing internal/race from internal/runtime/* packages, similar to internal/asan and internal/msan. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: Ibd9644557782076e3cee7927c8a6e6d2909f0a6e Reviewed-on: https://go-review.googlesource.com/c/go/+/616458 Reviewed-by: Keith Randall <khr@golang.org> Auto-Submit: Michael Pratt <mpratt@google.com> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Keith Randall <khr@google.com>
…ime/maps For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: Iebc7f5482299cb7c4ecccc4c2eb46b4bc42c5fc3 Reviewed-on: https://go-review.googlesource.com/c/go/+/616459 Reviewed-by: Keith Randall <khr@google.com> Auto-Submit: Michael Pratt <mpratt@google.com> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Keith Randall <khr@golang.org>
This is the same design as existing maps. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: I5f6ef5fea1e0f0616bcd90eaae7faee4cdac58c6 Reviewed-on: https://go-review.googlesource.com/c/go/+/616460 Reviewed-by: Keith Randall <khr@golang.org> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Auto-Submit: Michael Pratt <mpratt@google.com> Reviewed-by: Keith Randall <khr@google.com>
We use the same heuristics as existing maps. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: I44bb51483cae2c1714717f1b501850fb9e55a39a Reviewed-on: https://go-review.googlesource.com/c/go/+/616461 Auto-Submit: Michael Pratt <mpratt@google.com> Reviewed-by: Keith Randall <khr@google.com> Reviewed-by: Keith Randall <khr@golang.org> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Add all the specialized variants that exist for the existing maps. Like the existing maps, the fast variants do not support indirect key/elem. Note that as of this CL, the Get and Put methods on Map/table are effectively dead. They are only reachable from the internal/runtime/maps unit tests. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: I95297750be6200f34ec483e4cfc897f048c26db7 Reviewed-on: https://go-review.googlesource.com/c/go/+/616463 Reviewed-by: Keith Randall <khr@golang.org> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Auto-Submit: Michael Pratt <mpratt@google.com> Reviewed-by: Keith Randall <khr@google.com>
Keep only a single seed; initialize it; and reset it when the map is empty. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: Icc231f70957337a2d0dcd9c7daf9bd3cb4354d71 Reviewed-on: https://go-review.googlesource.com/c/go/+/616466 Auto-Submit: Michael Pratt <mpratt@google.com> Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Keith Randall <khr@google.com> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
This avoids some zero-extension ops on 64-bit machines. Based on khr@'s CL 619479. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: Ie9a56da26382dc9e515c613abc8cf6fec3767671 Reviewed-on: https://go-review.googlesource.com/c/go/+/620216 LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Keith Randall <khr@google.com> Auto-Submit: Michael Pratt <mpratt@google.com>
Load the field we need from the type once outside the search loop. Get rid of the multiply to compute the slot position. Instead compute the slot position incrementally using addition. Move the hashing later in access2. Based on khr@'s CL 618959. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: Id11b5479fa5bc0130a1d8d9e664d0206d24942ea Reviewed-on: https://go-review.googlesource.com/c/go/+/620217 Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Keith Randall <khr@google.com> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Auto-Submit: Michael Pratt <mpratt@google.com>
Speeds up iteration by about 3%. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: I3406376fb8db87306d52e665fcee1f33cf610f24 Reviewed-on: https://go-review.googlesource.com/c/go/+/621115 LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Keith Randall <khr@google.com> Auto-Submit: Michael Pratt <mpratt@google.com>
A previous CL kept it across loop iterations, but those are more rare than call iterations. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: Ieea0f1677e357f5e451650b1c697da7f63f3bca1 Reviewed-on: https://go-review.googlesource.com/c/go/+/621116 Reviewed-by: Keith Randall <khr@google.com> Auto-Submit: Michael Pratt <mpratt@google.com> Reviewed-by: Keith Randall <khr@golang.org> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
The compiler will stack allocate the Map struct and initial group if possible. Stack maps are initialized inline without calling into the runtime. Small heap allocated maps use makemap_small. These are the same heuristics as existing maps. For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-swissmap Change-Id: I6c371d1309716fd1c38a3212d417b3c76db5c9b9 Reviewed-on: https://go-review.googlesource.com/c/go/+/622042 LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Keith Randall <khr@golang.org> Reviewed-by: Keith Randall <khr@google.com> Auto-Submit: Michael Pratt <mpratt@google.com>
For #54766. Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest,gotip-linux-amd64-longtest-race,gotip-linux-arm64-longtest,gotip-linux-386-longtest,gotip-darwin-amd64-longtest,gotip-darwin-arm64_13,gotip-linux-ppc64_power10,gotip-linux-arm Change-Id: I5db0edcc156ed2e4bedc036b0baba2669e10c87a Reviewed-on: https://go-review.googlesource.com/c/go/+/594597 Reviewed-by: Keith Randall <khr@golang.org> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Keith Randall <khr@google.com>
|
Extendible hashing splitting can be implemented such that the bucket to split is kept around as the other half with 50% of the entries just moved to the other bucket that we allocate. This means that we end up allocating only half as much when splitting without creating any garbage. My idea: use arena allocator for the buckets because growing no longer introduces fragments of unused memory into the arena. I experimented with it here. The thing that you guys are cooking up right now may end up being a disappointment but integrating it with the experimental |
|
Yes, it would be nice if we could use the old table as half of the split. |
I think this is the hard part. Anything that removes keys from groups (when they haven't actually been deleted) makes iteration difficult to support. This same problem is what made me drop "rehash in place" (in place removal of tombstones). If we change the contents of the groups how do we know which keys we've already iterated over? |
|
@randall77 |
|
After a quick read through the implementation in internal/runtime/maps I found very little to complain about. But! The globalShift := 64 - globalDepth
entries := 1 << (globalDepth-table.localDepth)
index := (hash>>globalShift) &^ (entries - 1)
// index == table.index
// ... then after splitting
replaceTable(index , entries/2, left)
replaceTable(index + entries/2, entries/2, right)Also consider adding a comment on what happens if during splitting all entries end up being moved into just one table. Very very very unlikely thing to happen but it's clearly an edge-case so documenting why it's OK is likely a good idea. |
I agree that it is probably not strictly necessary. The uses:
While looking at this I also realized that |
While this is true, this issue is focused on the builtin map type where the language spec constrains iteration semantics. Other implementations can and do have different semantics (e.g., |
|
Change https://go.dev/cl/625517 mentions this issue: |
Now that GOEXPERIMENT=swissmap is enabled by default, switch to a -noswissmap builder to ensure the old maps continue to work until the GOEXPERIMENT is deleted. For golang/go#54766. Change-Id: I47cd081cd144740ccb95352b78d3a292b3e2a418 Reviewed-on: https://go-review.googlesource.com/c/build/+/625517 Reviewed-by: Dmitri Shuralyov <dmitshur@golang.org> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
|
Change https://go.dev/cl/626277 mentions this issue: |
|
Change https://go.dev/cl/626755 mentions this issue: |
CL 622042 added rand as a compiler builtin, but did not update builtinlist. Also update the mkbuiltin comment to refer to the current file location, and add a comment for runtime.rand that it is called from the compiler. For #54766 Change-Id: I99d2c0bb0658da333775afe2ed0447265c845c82 Reviewed-on: https://go-review.googlesource.com/c/go/+/626755 Reviewed-by: Cherry Mui <cherryyz@google.com> Reviewed-by: Michael Pratt <mpratt@google.com> LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Auto-Submit: Ian Lance Taylor <iant@golang.org>
|
Change https://go.dev/cl/627156 mentions this issue: |
Iteration over swissmaps with low load (think map with large hint but
only one entry) is signicantly regressed vs old maps. See noswiss vs
swiss-tip below (+60%).
Currently we visit every single slot and individually check if the slot
is full or not.
We can do much better by using the control word to find all full slots
in a group in a single operation. This lets us skip completely empty
groups for instance.
Always using the control match approach is great for maps with low load,
but is a regression for mostly full maps. Mostly full maps have the
majority of slots full, so most calls to mapiternext will return the
next slot. In that case, doing the full group match on every call is
more expensive than checking the individual slot.
Thus we take a hybrid approach: on each call, we first check an
individual slot. If that slot is full, we're done. If that slot is
non-full, then we fall back to doing full group matches.
This trade-off works well. Both mostly empty and mostly full maps
perform nearly as well as doing all matching and all individual,
respectively.
The fast path is placed above the slow path loop rather than combined
(with some sort of `useMatch` variable) into a single loop to help the
compiler's code generation. The compiler really struggles with code
generation on a combined loop for some reason, yielding ~15% additional
instructions/op.
Comparison with old maps prior to this CL:
│ noswiss │ swiss-tip │
│ sec/op │ sec/op vs base │
MapIter/Key=int64/Elem=int64/len=6-12 11.53n ± 2% 10.64n ± 2% -7.72% (p=0.002 n=6)
MapIter/Key=int64/Elem=int64/len=64-12 10.180n ± 2% 9.670n ± 5% -5.01% (p=0.004 n=6)
MapIter/Key=int64/Elem=int64/len=65536-12 10.78n ± 1% 10.15n ± 2% -5.84% (p=0.002 n=6)
MapIterLowLoad/Key=int64/Elem=int64/len=6-12 6.116n ± 2% 6.840n ± 2% +11.84% (p=0.002 n=6)
MapIterLowLoad/Key=int64/Elem=int64/len=64-12 2.403n ± 2% 3.892n ± 0% +61.95% (p=0.002 n=6)
MapIterLowLoad/Key=int64/Elem=int64/len=65536-12 1.940n ± 3% 3.237n ± 1% +66.81% (p=0.002 n=6)
MapPop/Key=int64/Elem=int64/len=6-12 66.20n ± 2% 60.14n ± 3% -9.15% (p=0.002 n=6)
MapPop/Key=int64/Elem=int64/len=64-12 97.24n ± 1% 171.35n ± 1% +76.21% (p=0.002 n=6)
MapPop/Key=int64/Elem=int64/len=65536-12 826.1n ± 12% 842.5n ± 10% ~ (p=0.937 n=6)
geomean 17.93n 20.96n +16.88%
After this CL:
│ noswiss │ swiss-cl │
│ sec/op │ sec/op vs base │
MapIter/Key=int64/Elem=int64/len=6-12 11.53n ± 2% 10.90n ± 3% -5.42% (p=0.002 n=6)
MapIter/Key=int64/Elem=int64/len=64-12 10.180n ± 2% 9.719n ± 9% -4.53% (p=0.043 n=6)
MapIter/Key=int64/Elem=int64/len=65536-12 10.78n ± 1% 10.07n ± 2% -6.63% (p=0.002 n=6)
MapIterLowLoad/Key=int64/Elem=int64/len=6-12 6.116n ± 2% 7.022n ± 1% +14.82% (p=0.002 n=6)
MapIterLowLoad/Key=int64/Elem=int64/len=64-12 2.403n ± 2% 1.475n ± 1% -38.63% (p=0.002 n=6)
MapIterLowLoad/Key=int64/Elem=int64/len=65536-12 1.940n ± 3% 1.210n ± 6% -37.67% (p=0.002 n=6)
MapPop/Key=int64/Elem=int64/len=6-12 66.20n ± 2% 61.54n ± 2% -7.02% (p=0.002 n=6)
MapPop/Key=int64/Elem=int64/len=64-12 97.24n ± 1% 110.10n ± 1% +13.23% (p=0.002 n=6)
MapPop/Key=int64/Elem=int64/len=65536-12 826.1n ± 12% 504.7n ± 6% -38.91% (p=0.002 n=6)
geomean 17.93n 15.29n -14.74%
For #54766.
Cq-Include-Trybots: luci.golang.try:gotip-linux-ppc64_power10
Change-Id: Ic07f9df763239e85be57873103df5007144fdaef
Reviewed-on: https://go-review.googlesource.com/c/go/+/627156
Auto-Submit: Michael Pratt <mpratt@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Keith Randall <khr@google.com>
Abstract
We suggest using SwissTable in the runtime to replace the original implementation of the hashmap.
SwissTable is now used in abseil and rust by default. We borrowed ideas from the two implementations and combined some optimization ideas of Go's original hashmap to construct the current implementation.
See the following comment for performance comparison(the GitHub issue has a character limit).
Motivations
We need to improve the performance of the hashmap as much as possible. At ByteDance, our Go service consumes about 4% of the CPU on hashmap. It is worth noting that on many services, the CPU consumption of
mapassignandmapaccessare almost 1:1, so the improvement of insertion performance is equally important.Support dynamic adjustment of the capacity of the hashmap. Some Go services will make a map too large due to burst traffic, which will cause a lot of pressure on the memory, but the map does not shrink after elements removal. We found that there are also many related discussions in the community:
Implementations
Basic version: https://go-review.googlesource.com/c/go/+/426614
SIMD version: TODO
Tasks
Advantages compared to the original implementation
The overall performance is better, and we can also use SIMD to improve performance.
SwissTable's LoadFactor can be set higher to save some memory.
Open-addressing and making dynamic adjustments to the capacity is easier to implement.
After allocating a fixed-size hashmap, if the number of inserted elements does not exceed the capacity, no additional memory is required, and the performance will be significantly improved, which has a greater advantage over reused fixed-size hashmap. (The original hashmap may still allocate additional memory even when the limit is not exceeded)
Disadvantages compared to the original implementation
The rehash is done at once, so
Since SwissTable needs to ensure that each bucket has at least one empty slot, it will cause performance degradation in some cases. For example, if we insert an element when there are seven elements in the hashmap, the CPU consumption will significantly increase because the hashmap needs to grow.
Implementation details
Here is an overview of the basic version. For more detailed implementation, you can read the code. Before reading the code, it is recommended to check the video or the article in the references to get a general idea of how it works.
Differences with SwissTable
tophashinto the bucket, not as an array alone.The classic SwissTable aggregates all tophash into an array (also called control bytes array, here we follow the convention, called tophash), while the original hashmap of Go puts tophash and items in the same bucket.
Our experiments outside the runtime found that if tophash is used as an array alone, we couldn’t find any performance wins. This is probably because we need two memory allocations when initializing the hashmap.
So we use the bucket memory layout almost the same as the original in the current version. In the future, we may consider aggregating all tophash into an array, using a data structure similar to
sync.Pool, so that all tophash of the hashmap can be reused to improve performance.Differences with the original implementation
For hashmap header (
hmap) andhiterFor bucket (
bmap)The tophash is still 8-bit(uint8), but 7-bit for hash info and 1-bit as flags.
Removed overflow pointer.
For the way of querying
For initializing bucket arrays
0xffinstead of0;0xffrepresents the empty slot now.References
C++: https://github.com/abseil/abseil-cpp/blob/master/absl/container/internal/raw_hash_set.h
Rust: https://github.com/rust-lang/hashbrown
The text was updated successfully, but these errors were encountered: