switch gc for collection benchmarks

[github-actions[bot]]

Note Diffing the performance result against the published result from main branch. Unchanged benchmarks are omitted.

Map

	binary_size	generate 1m	max mem	batch_get 50	batch_put 50	batch_remove 50
hashmap	136_413 ($\textcolor{red}{1.93\%}$)	7_606_067_149 ($\textcolor{red}{9.28\%}$)	61_987_708	288_199 ($\textcolor{red}{0.25\%}$)	6_162_750_199 ($\textcolor{red}{11.73\%}$)	309_738 ($\textcolor{red}{0.25\%}$)
triemap	137_903 ($\textcolor{red}{1.91\%}$)	12_189_110_689 ($\textcolor{red}{6.63\%}$)	74_215_992	222_822 ($\textcolor{red}{0.02\%}$)	547_698 ($\textcolor{red}{0.01\%}$)	539_062 ($\textcolor{red}{0.01\%}$)
rbtree	138_702 ($\textcolor{red}{1.90\%}$)	6_610_828_306 ($\textcolor{red}{10.56\%}$)	57_995_880	88_902 ($\textcolor{red}{0.00\%}$)	268_570 ($\textcolor{red}{0.00\%}$)	278_336 ($\textcolor{red}{0.00\%}$)
splay	134_438 ($\textcolor{red}{1.95\%}$)	12_359_483_042 ($\textcolor{red}{6.84\%}$)	53_995_816	551_923 ($\textcolor{red}{0.00\%}$)	581_661 ($\textcolor{red}{0.00\%}$)	810_217 ($\textcolor{red}{0.00\%}$)
btree	178_981 ($\textcolor{red}{1.43\%}$)	8_510_551_526 ($\textcolor{red}{3.48\%}$)	31_103_868	277_539 ($\textcolor{red}{0.00\%}$)	384_168 ($\textcolor{red}{0.00\%}$)	429_038 ($\textcolor{red}{0.00\%}$)
zhenya_hashmap	144_414 ($\textcolor{red}{1.80\%}$)	3_460_663_105 ($\textcolor{red}{25.56\%}$)	65_987_456	68_446 ($\textcolor{red}{0.08\%}$)	83_148 ($\textcolor{red}{0.06\%}$)	93_334 ($\textcolor{red}{0.07\%}$)
btreemap_rs	413_478	1_649_709_879	13_762_560	66_814	112_263	81_263
imrc_hashmap_rs	413_588	2_385_702_121	122_454_016	32_846	162_715	98_494
hashmap_rs	406_096	392_593_368	36_536_320	16_498	20_863	19_973

Priority queue

	binary_size	heapify 1m	max mem	pop_min 50	put 50
heap	130_340 ($\textcolor{red}{2.03\%}$)	4_972_973_442 ($\textcolor{red}{6.16\%}$)	29_995_812	511_496 ($\textcolor{red}{0.00\%}$)	186_462 ($\textcolor{red}{0.00\%}$)
heap_rs	403_925	123_102_482	9_109_504	53_320	18_138

Statistics

• binary_size: 1.85% [1.71%, 2.00%]
• max_mem: no change
• cycles: 3.00% [1.08%, 4.92%]

Basic DAO

	binary_size	init	transfer_token	submit_proposal	vote_proposal
Motoko	225_805	37_493 ($\textcolor{red}{0.06\%}$)	16_270 ($\textcolor{red}{0.26\%}$)	12_654 ($\textcolor{green}{-0.38\%}$)	14_126 ($\textcolor{green}{-0.21\%}$)
Rust	704_886	471_865	86_470	104_617	115_765

DIP721 NFT

Note Same as main branch, skipping.

Statistics

• binary_size: no change
• max_mem: no change
• cycles: -0.07% [-0.40%, 0.27%]

Heartbeat

	binary_size	heartbeat
Motoko	118_909	7_392
Rust	23_699	474 ($\textcolor{green}{-40.08\%}$)

Timer

Note Same as main branch, skipping.

Statistics

• binary_size: no change

• max_mem: no change

• cycles: no change

  Overall Statistics

• binary_size: 1.85% [1.71%, 2.00%]

• max_mem: no change

• cycles: 2.60% [0.92%, 4.29%]

[github-actions[bot]]

Note The flamegraph link only works after you merge. Unchanged benchmarks are omitted.

Collection libraries

Measure different collection libraries written in both Motoko and Rust. The library names with _rs suffix are written in Rust; the rest are written in Motoko.

We use the same random number generator with fixed seed to ensure that all collections contain the same elements, and the queries are exactly the same. Below we explain the measurements of each column in the table:

• generate 1m. Insert 1m Nat64 integers into the collection. For Motoko collections, it usually triggers the GC; the rest of the column are not likely to trigger GC.
• max mem. For Motoko, it reports rts_max_live_size after generate call; For Rust, it reports the Wasm's memory page * 32Kb.
• batch_get 50. Find 50 elements from the collection.
• batch_put 50. Insert 50 elements to the collection.
• batch_remove 50. Remove 50 elements from the collection.

💎 Takeaways

• The platform only charges for instruction count. Data structures which make use of caching and locality have no impact on the cost.
• We have a limit on the maximal cycles per round. This means asymptotic behavior doesn't matter much. We care more about the performance up to a fixed N. In the extreme cases, you may see an O(10000 nlogn) algorithm hitting the limit, while an O(n^2) algorithm runs just fine.
• Amortized algorithms/GC may need to be more eager to avoid hitting the cycle limit on a particular round.
• Rust costs more cycles to process complicated Candid data, but it is more efficient in performing core computations.

Note

• The Candid interface of the benchmark is minimal, therefore the serialization cost is negligible in this measurement.
• Due to the instrumentation overhead and cycle limit, we cannot profile computations with large collections. Hopefully, when deterministic time slicing is ready, we can measure the performance on larger memory footprint.
• hashmap uses amortized data structure. When the initial capacity is reached, it has to copy the whole array, thus the cost of batch_put 50 is much higher than other data structures.
• btree comes from Byron Becker's stable BTreeMap library.
• zhenya_hashmap comes from Zhenya Usenko's stable HashMap library.
• hashmap_rs uses the fxhash crate, which is the same as std::collections::HashMap, but with a deterministic hasher. This ensures reproducible result.
• imrc_hashmap_rs uses the im-rc crate, which is the immutable version hashmap in Rust.
• The MoVM table measures the performance of an experimental implementation of Motoko interpreter. External developers can ignore this table for now.

Map

	binary_size	generate 1m	max mem	batch_get 50	batch_put 50	batch_remove 50
hashmap	136_413	7_606_067_149	61_987_708	288_199	6_162_750_199	309_738
triemap	137_903	12_189_110_689	74_215_992	222_822	547_698	539_062
rbtree	138_702	6_610_828_306	57_995_880	88_902	268_570	278_336
splay	134_438	12_359_483_042	53_995_816	551_923	581_661	810_217
btree	178_981	8_510_551_526	31_103_868	277_539	384_168	429_038
zhenya_hashmap	144_414	3_460_663_105	65_987_456	68_446	83_148	93_334
btreemap_rs	413_478	1_649_709_879	13_762_560	66_814	112_263	81_263
imrc_hashmap_rs	413_588	2_385_702_121	122_454_016	32_846	162_715	98_494
hashmap_rs	406_096	392_593_368	36_536_320	16_498	20_863	19_973

Priority queue

	binary_size	heapify 1m	max mem	pop_min 50	put 50
heap	130_340	4_972_973_442	29_995_812	511_496	186_462	487_203
heap_rs	403_925	123_102_482	9_109_504	53_320	18_138	53_543

Sample Dapps

Measure the performance of some typical dapps:

• Basic DAO, with heartbeat disabled to make profiling easier. We have a separate benchmark to measure heartbeat performance.
• DIP721 NFT

Note

• The cost difference is mainly due to the Candid serialization cost.
• Motoko statically compiles/specializes the serialization code for each method, whereas in Rust, we use serde to dynamically deserialize data based on data on the wire.
• We could improve the performance on the Rust side by using parser combinators. But it is a challenge to maintain the ergonomics provided by serde.
• For real-world applications, we tend to send small data for each endpoint, which makes the Candid overhead in Rust tolerable.

Basic DAO

	binary_size	init	transfer_token	submit_proposal	vote_proposal
Motoko	225_805	37_493	16_270	12_654	14_126
Rust	704_886	471_865	86_470	104_617	115_765

DIP721 NFT

	binary_size	init	mint_token	transfer_token
Motoko	183_882	12_181	22_319	4_710
Rust	766_710	125_034	324_482	77_116

Heartbeat / Timer

Measure the cost of empty heartbeat and timer job.

• setTimer measures both the setTimer(0) method and the execution of empty job.
• It is not easy to reliably capture the above events in one flamegraph, as the implementation detail of the replica can affect how we measure this. Typically, a correct flamegraph contains both setTimer and canister_global_timer function. If it's not there, we may need to adjust the script.

Heartbeat

	binary_size	heartbeat
Motoko	118_909	7_392
Rust	23_699	474

Timer

	binary_size	setTimer	cancelTimer
Motoko	125_168	15_208	1_679
Rust	434_848	43_540	7_683