They've made SmallRng have a different length seed on different platforms, which is .. an odd choice. Instead of messing around with generating a seed, I've taken their recommendation and switched to a specific implementation (the one SmallRng uses in 0.8 anyway), which uses a bigger seed.
Follows #205 so the CI will work.
With lto and RUSTFLAGS='-C target-cpu=native' on my i7-8700K:
Running `target/release/perftest 10000000`
labels rust-protobuf quick-protobuf prost quick/rust prost/rust
ns/iter ns/iter ns/iter % %
test1
write 35 10 9 71.4 74.3
read 28 22 16 21.4 42.9
read no vec 20 17 12 15.0 40.0
read reuse 20 NA NA
test_repeated_bool
write 40 13 10 67.5 75.0
read 67 53 53 20.9 20.9
read no vec 43 35 34 18.6 20.9
read reuse 29 NA NA
test_repeated_packed_int32
write 55 27 21 50.9 61.8
read 89 79 70 11.2 21.3
read no vec 63 58 51 7.9 19.0
read reuse 34 NA NA
test_repeated_messages
write 125 80 115 36.0 8.0
read 666 509 494 23.6 25.8
read no vec 186 167 144 10.2 22.6
read reuse 92 NA NA
test_optional_messages
write 111 56 74 49.5 33.3
read 250 143 130 42.8 48.0
read no vec 143 115 96 19.6 32.9
read reuse 86 NA NA
test_strings
write 56 25 22 55.4 60.7
read 134 69 155 48.5 -15.7
read no vec 93 48 124 48.4 -33.3
read reuse 59 NA NA
test_small_bytearrays
write 215 158 189 26.5 12.1
read 255 103 291 59.6 -14.1
read no vec 111 31 142 72.1 -27.9
read reuse 103 NA NA
test_large_bytearrays
write 14141 13895 13575 1.7 4.0
read 14018 4557 16353 67.5 -16.7
read no vec 6246 46 7499 99.3 -20.1
read reuse 7246 NA NA
Also (not performance related but fun to know) - Rpc test was successful!
Upgrade
randin perftest.They've made
SmallRnghave a different length seed on different platforms, which is .. an odd choice. Instead of messing around with generating a seed, I've taken their recommendation and switched to a specific implementation (the one SmallRng uses in 0.8 anyway), which uses a bigger seed.Follows #205 so the CI will work.
With
ltoandRUSTFLAGS='-C target-cpu=native'on my i7-8700K: