Add unsafe blake3 benchmark for merkledb

[StephenButtolph]

Why this should be merged

Built on top of the safe blake3 benchmark: #3349.

This PR isn't safe, as the hasher interface expects to be able to be used concurrently. However, it shows the potential performance of BLAKE3 against SHA256.

goos: darwin
goarch: arm64
pkg: github.com/ava-labs/avalanchego/x/merkledb
Benchmark_SHA256_HashNode/empty_node-12         	14426436	        79.50 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_value-12          	14245774	        84.81 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_key-12            	16405936	        72.80 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/1_child-12            	 8694835	       138.8 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/2_children-12         	 5892787	       201.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/16_children-12        	 1000000	      1052 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/empty_node-12         	 9002222	       131.4 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_value-12          	 8012547	       149.9 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_key-12            	 8915469	       136.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/1_child-12            	 6148550	       195.5 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/2_children-12         	 3697560	       324.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/16_children-12        	  751521	      1606 ns/op	       0 B/op	       0 allocs/op

^ on my macbook pro (apple M2 Max chip)

How this works

Resetting the BLAKE3 internal state is significantly faster than initializing an entirely new hasher.

How this was tested

• Existing unit tests + benchmarks

[StephenButtolph]

Running on a new Ubuntu AWS instance using c7i.large (amd64) gives:

goos: linux
goarch: amd64
pkg: github.com/ava-labs/avalanchego/x/merkledb
cpu: Intel(R) Xeon(R) Platinum 8488C
Benchmark_SHA256_HashNode/empty_node-2         	11140938	       107.4 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_value-2          	10032351	       119.9 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_key-2            	10851236	       115.2 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/1_child-2            	 6390860	       182.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/2_children-2         	 4471764	       267.4 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/16_children-2        	  891460	      1319 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/empty_node-2         	11008384	       108.9 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_value-2          	10042704	       120.2 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_key-2            	10968001	       112.7 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/1_child-2            	 6379234	       183.3 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/2_children-2         	 4328890	       279.5 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/16_children-2        	  798664	      1484 ns/op	       0 B/op	       0 allocs/op

Which is much more similar between the two implementations (SHA256 still being slightly faster)

The performance is similar when running on c7a.large.

[StephenButtolph]

Running on a new Ubuntu AWS instance using c7g.large (arm64) gives:

goos: linux
goarch: arm64
pkg: github.com/ava-labs/avalanchego/x/merkledb
Benchmark_SHA256_HashNode/empty_node-2         	 8024362	       149.3 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_value-2          	 6909513	       173.0 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_key-2            	 8041004	       149.4 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/1_child-2            	 4334109	       276.4 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/2_children-2         	 3111552	       385.4 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/16_children-2        	  656708	      1861 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/empty_node-2         	 5403916	       222.2 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_value-2          	 4870874	       246.3 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_key-2            	 5143668	       233.2 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/1_child-2            	 3152232	       381.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/2_children-2         	 2070284	       579.5 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/16_children-2        	  426556	      2795 ns/op	       0 B/op	       0 allocs/op

[StephenButtolph]

Running on a new Ubuntu AWS instance using c6i.large (amd64) gives:

goos: linux
goarch: amd64
pkg: github.com/ava-labs/avalanchego/x/merkledb
cpu: Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz
Benchmark_SHA256_HashNode/empty_node-2         	 9677665	       124.2 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_value-2          	 8577712	       140.2 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_key-2            	 9545270	       125.9 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/1_child-2            	 5190390	       232.5 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/2_children-2         	 3504306	       342.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/16_children-2        	  714806	      1661 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/empty_node-2         	10049533	       119.8 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_value-2          	 8944796	       136.6 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_key-2            	 9560641	       126.3 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/1_child-2            	 5372684	       223.5 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/2_children-2         	 3504469	       342.7 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/16_children-2        	  618915	      1723 ns/op	       0 B/op	       0 allocs/op

Interestingly - for the smallest nodes, BLAKE3 actually outperforms here (yet loses against the "bigger" sizes - which are still very small)

[StephenButtolph]

Running on a new Ubuntu AWS instance using c6g.large (arm64) gives:

goos: linux
goarch: arm64
pkg: github.com/ava-labs/avalanchego/x/merkledb
Benchmark_SHA256_HashNode/empty_node-2         	 6432463	       186.9 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_value-2          	 5515927	       217.4 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_key-2            	 6405484	       187.4 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/1_child-2            	 3292816	       362.7 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/2_children-2         	 2384853	       502.8 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/16_children-2        	  450978	      2732 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/empty_node-2         	 3864762	       310.5 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_value-2          	 3517705	       341.3 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_key-2            	 3777644	       317.6 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/1_child-2            	 2455851	       488.4 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/2_children-2         	 1535839	       781.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/16_children-2        	  302064	      3967 ns/op	       0 B/op	       0 allocs/op

[StephenButtolph]

Running on a new Ubuntu AWS instance using c5.large (amd64) gives:

goos: linux
goarch: amd64
pkg: github.com/ava-labs/avalanchego/x/merkledb
cpu: Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz
Benchmark_SHA256_HashNode/empty_node-2         	 4593003	       260.8 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_value-2          	 4272609	       280.6 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_key-2            	 4595166	       266.6 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/1_child-2            	 3111033	       385.6 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/2_children-2         	 1916672	       625.5 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/16_children-2        	  392150	      3029 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/empty_node-2         	 9074803	       132.3 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_value-2          	 7937150	       151.2 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_key-2            	 8880096	       135.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/1_child-2            	 4736965	       252.8 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/2_children-2         	 3229576	       371.5 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/16_children-2        	  630412	      1910 ns/op	       0 B/op	       0 allocs/op

This is the first instance that shows a significant performance improvement when using BLAKE3. Overall still slower than the c7i.large series though.

[StephenButtolph]

Running on a new Ubuntu AWS instance using c4.large (amd64) gives:

goos: linux
goarch: amd64
pkg: github.com/ava-labs/avalanchego/x/merkledb
cpu: Intel(R) Xeon(R) CPU E5-2666 v3 @ 2.90GHz
Benchmark_SHA256_HashNode/empty_node-2         	 4565466	       262.2 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_value-2          	 4233136	       282.7 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/has_key-2            	 4548741	       263.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/1_child-2            	 3088939	       388.7 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/2_children-2         	 1899824	       630.9 ns/op	       0 B/op	       0 allocs/op
Benchmark_SHA256_HashNode/16_children-2        	  393420	      3046 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/empty_node-2         	 9008563	       133.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_value-2          	 7875678	       152.2 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/has_key-2            	 8815430	       136.1 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/1_child-2            	 4703169	       255.6 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/2_children-2         	 3191654	       375.9 ns/op	       0 B/op	       0 allocs/op
Benchmark_BLAKE3_HashNode/16_children-2        	  613514	      1925 ns/op	       0 B/op	       0 allocs/op

Again BLAKE3 seems to outperform on these older chips.

[github-actions]

This PR has become stale because it has been open for 30 days with no activity. Adding the lifecycle/frozen label will cause this PR to ignore lifecycle events.