Can you add the implementation and example of wazero? about proxy-wasm-go-host

[antJack]

I don't think cgo has any problems.

As for complexity, cgo only appears under the hood. The default wasm runtime in our project is wasmer-go, so that users do not need to write any cgo codes by themselves.

As for performance, in our env, a go->c cgo call cost only ~60ns, it does much slower that nativa go->go/c->c call, but it can be ignored in proxy case, which usually cost ~ms for one request to be done.
Moreover, the wazero project do not provide any benchmark data, such as pure go wasm runtime vs cgo wasm runtime, so it can not convince me about their so-called performance.

If you insist on supporting the wazero, welcome to open up a PR to integrate the wazero runtime, just like what we do in https://github.com/mosn/proxy-wasm-go-host/tree/main/wasmer, it not a difficult task :-D

[antJack]

After a quick glance of the wazero project, I find that it's just another wasm runtime, just like V8, Wasmtime, Wasmer, WAVM, etc. The main difference is that the wazero is implemented in pure Go, so that it can get rid of cgo. However, its support on wasm core specification is not as good as other runtimes, such as Wasmer. So do we have any good reason to support the wazero runtime?

[pretty66]

it can get rid of cgo

[antJack]

Does cgo bring any problems?

[pretty66]

The main advantage or feature is this

https://github.com/tetratelabs/wazero#history-of-wazero

[codefromthecrypt]

ps I noticed this issue, and wazero does have benchmarks, just we let people run them on own. For example, our build breaks if the benchmarks are ever slower than wasmtime wasmer or wasm3.

ex they are run on every commit. Not sure how many others do that

https://github.com/tetratelabs/wazero/runs/6094956125?check_suite_focus=true

[codefromthecrypt]

Ex you can run make bench.check if you like, or execute the vs directory yourself. Here's an example from the above CI run.

goos: linux
goarch: amd64
pkg: github.com/tetratelabs/wazero/internal/integration_test/vs
cpu: Intel(R) Xeon(R) Platinum 8272CL CPU @ 2.60GHz
BenchmarkFac_Init/Interpreter-2         	   53996	     21683 ns/op	   11960 B/op	     370 allocs/op
BenchmarkFac_Init/JIT-2                 	    8391	    151112 ns/op	   62045 B/op	    1251 allocs/op
BenchmarkFac_Init/wasmer-go-2           	    1999	    598627 ns/op	    1232 B/op	      41 allocs/op
BenchmarkFac_Init/wasmtime-go-2         	    1938	    622997 ns/op	     345 B/op	      21 allocs/op
BenchmarkFac_Init/go-wasm3-2            	  137173	      8501 ns/op	     168 B/op	      11 allocs/op
BenchmarkFac_Invoke/Interpreter-2       	  118746	      9963 ns/op	    2096 B/op	     129 allocs/op
BenchmarkFac_Invoke/JIT-2               	  806187	      1349 ns/op	    1216 B/op	       5 allocs/op
BenchmarkFac_Invoke/wasmer-go-2         	  730215	      1688 ns/op	     176 B/op	      12 allocs/op
BenchmarkFac_Invoke/wasmtime-go-2       	  354572	      3711 ns/op	     256 B/op	      20 allocs/op
BenchmarkFac_Invoke/go-wasm3-2          	  484650	      2504 ns/op	      56 B/op	       5 allocs/op
BenchmarkCodecExample/binary.DecodeModule-2         	  233971	      5060 ns/op	    2440 B/op	     109 allocs/op
BenchmarkCodecExample/binary.EncodeModule-2         	  240770	      4950 ns/op	    3960 B/op	     124 allocs/op
BenchmarkCodecExample/text.DecodeModule-2           	   50046	     24196 ns/op	    8560 B/op	     369 allocs/op
BenchmarkCodecExample/wat2wasm_via_text.DecodeModule->binary.EncodeModule-2         	   40120	     29765 ns/op	   12528 B/op	     493 allocs/op
BenchmarkCodecExample/wat2wasm_vs_wasmer.Wat2Wasm-2                                 	   24598	     48232 ns/op	    2704 B/op	       6 allocs/op
BenchmarkCodecExample/wat2wasm_vs_wasmtime.Wat2Wasm-2                               	   29388	     40440 ns/op	    2672 B/op	       3 allocs/op
PASS
ok  	github.com/tetratelabs/wazero/internal/integration_test/vs	23.480s

In any case, I think it is a better fit for folks who desire specific things that are impossible otherwise. Ex multi-platform builds, transparent debug story (ex no clib blackhole), coherent trace context propagation etc. If these things aren't relevant and you're otherwise happy where you are, maybe best to stick with it. Meanwhile, good luck wherever you end up!

[codefromthecrypt]

ps also maybe not relevant to you but image size in proxies is sometimes a problem. Wazero can be 3MB total in docker, and I think that's pretty impossible in any of the alternatives tetratelabs/wazero#424

[codefromthecrypt]

one last spam. I learned a lot reading this on CGO it does a great walk through of things in a different way than the usual "CGO is not go" blogs. Hope it helps you decide not so much if wasmtime should be in here, but more should CGO be the only way to use wasm in the project. https://speakerdeck.com/jcchavezs/cgo-journey-from-using-to-understanding-it

PS it is ironic this issue exists because months ago I thought exactly mosn would be a good thing to try to work in wazero, especially for unit testing proxy plugins without any hassle or weird windows things etc. just thought it would be cheesy to make an issue to suggest it, being that I'm a maintainer. I'm glad someone else raised this, so I had an excuse to spam you all these comments. But don't worry, that's all unless you ask for more 😎

[codefromthecrypt]

#19 but I think we can do a lot better as noted at the end of my description. The attempt to abstract things forces things to require wrapping and reflection, the performance impact is really bad and also the code is harder to follow vs doing things directly. Later I'd be open to rewriting it to be direct and fast if you like.

[codefromthecrypt]

#27 will add this. we've figured our way through a misunderstanding. Notably, it is ok to fix the v2 and also wazero made a concession to allow the unsupported wasi_unstable to work as some old code may never change.

[codefromthecrypt]

@pretty66 once mosn/mosn#2172 is merged "wazero" is a valid engine choice in mosn config!

[codefromthecrypt]

@pretty66 v1.3.0 mosn is out and includes this now. thanks for your patience and all the support from @antJack and @taoyuanyuan!