Skip to content

guillaumebort/mison

Repository files navigation

Experiments for a JVM Mison implementation

See Mison paper at: https://www.microsoft.com/en-us/research/wp-content/uploads/2017/05/mison-vldb17.pdf

Implementation detail

It follows exactly the design described in the Mison paper. Because there is no way to use avx2 extensions from the JVM this part is written in rust and wrapped in a native call. Also the predictive parsing part is not implemented, but because of the cost of building the Mison indexes I don't think it will change anything to the benchmark for small Json documents.

Benchmarks

Benchmark            (entry)  Mode  Cnt       Score       Error  Units

Benchmarks.fastJson    empty  avgt  200     244.944 ±     25.159  ns/op
Benchmarks.fastJson    short  avgt  200    1171.627 ±     51.561  ns/op
Benchmarks.fastJson   medium  avgt  200    2967.922 ±    156.079  ns/op
Benchmarks.fastJson     long  avgt  200   12229.239 ±    407.896  ns/op
Benchmarks.fastJson     huge  avgt  200  118164.704 ±   3499.133  ns/op
Benchmarks.fastJson    giant  avgt  200  438202.179 ±  19309.856  ns/op

Benchmarks.jackson     empty  avgt  200    4872.127 ±   2271.468  ns/op
Benchmarks.jackson     short  avgt  200    9131.281 ±   7338.789  ns/op
Benchmarks.jackson    medium  avgt  200    5470.434 ±    534.308  ns/op
Benchmarks.jackson      long  avgt  200   11512.887 ±    889.058  ns/op
Benchmarks.jackson      huge  avgt  200  569606.131 ± 997983.165  ns/op
Benchmarks.jackson     giant  avgt  200  517020.023 ± 742322.271  ns/op

Benchmarks.mison       empty  avgt  200    2033.237 ±     85.180  ns/op
Benchmarks.mison       short  avgt  200    3312.618 ±    141.704  ns/op
Benchmarks.mison      medium  avgt  200    5707.338 ±    264.357  ns/op
Benchmarks.mison        long  avgt  200    8287.681 ±    562.832  ns/op
Benchmarks.mison        huge  avgt  200   38999.464 ±   3946.958  ns/op
Benchmarks.mison       giant  avgt  200  137388.974 ±  13729.804  ns/op

Conclusion

This mison implementation starts being more performant at the long test case (around 1kb). The overhead of the native call makes it less interesting for small json.

Perhaps it is possible to reduce this overhead (for example by using offheap memory directly from JVM code, and then just sharing a raw pointer with the native code). But for now I'm not convinced that for the common use case it will be possible to do something better as something like FastJson on the JVM.

About

Scala Mison implementation

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published