benchmark

[yamt]

No description provided.

[Robbepop]

This can be emulated with Wasmi, too via --compilation-mode lazy-translation. Though I generally wouldn't recommend using that since it is just roughly 50% faster than eager translation but has all the downsides of lazy translation.

[yamt]

ok. i didn't know the option.
personally i like lazy-translation mode because it wouldn't change visible behaviors from eager.

[Robbepop]

Yes, I also like it a lot which is why I introduced both lazy variants. But in practical use cases it often happens that inputs are pre-validated so the lazy option is as safe and much more efficient than the lazy-translation option.

[Robbepop]

With respect to my comment about unhopeful inefficiencies during runtime it seems Wasmi got installed properly by the script with all needed optimizations. One thing that never occurred to me before was how important march=native could be on some of the hardware systems to avoid the mentioned inefficiencies in #8 (comment).

[Robbepop]

What I mean is Cargo's -C target-cpu=native option. At least according to this article (https://vfoley.xyz/rust-compilation-tip/) it can sometimes help. E.g. via RUSTFLAGS="-C target-cpu=native"; cargo +stable build --profile bench -p wasmi_cli.

[yamt]

interesting. i might try it when i have a chance.
however, for those runtimes i built myself here, i didn't apply such optimizations to any of them. (unless it's the default, which i haven't investigated.)
do you think wasmi is more sensitive to this kind of optimizations than other runtimes?

[Robbepop]

I wish I had a machine where Wasmi runs badly on so I could test it on my own. I wouldn't want Wasmi to have special optimization flags that no other runtime needs, rather I would like to understand the reason and be able to make Wasmi optimizations more robust if only I knew what was causing the issues in the first place. 🙃

do you think wasmi is more sensitive to this kind of optimizations than other runtimes?

Yes, I think a big reason for this is the missing control in the engine's instruction dispatch (hot path) due to missing control over the codegen with only a loop+match construct instead of something more elaborate such as tail call dispatching. Unfortunately Rust does not support tail calls or any other usable form of dispatching such as computed goto etc.
So the best we can do today in Wasmi is to pray that the LLVM optimizer heuristics get the instruction dispatch codegen right.

[yamt]

RUSTFLAGS="-C target-cpu=native"; cargo +stable build --profile bench -p wasmi_cli

i tried this. i haven't noticed any improvements.

spacetanuki% ./test/run-ffmpeg.sh /usr/bin/time -l wasmi_cli.target-cpu-native -
-dir .video --                                                                  
executing File(".ffmpeg/ffmpeg.wasm")::_start() ... 
       32.14 real        32.08 user         0.04 sys
           112017408  maximum resident set size      
                   0  average shared memory size 
                   0  average unshared data size                                
                   0  average unshared stack size                               
               30677  page reclaims                                             
                   0  page faults                                               
                   0  swaps                                                     
                   0  block input operations                                    
                   0  block output operations                                   
                   0  messages sent    
                   0  messages received                                         
                   0  signals received                                          
                   0  voluntary context switches  
                 660  involuntary context switches                              
        254158714732  instructions retired                                      
        129063009419  cycles elapsed                                            
            90820608  peak memory footprint                                     
spacetanuki% 

[Robbepop]

It's sad that the new Wasmi runs so poorly on your machine but I see it as my job to improve its performance on all platforms if that's possible for me. Thank you for taking all your time and feedback!

[Robbepop]

btw. if you rename this CVS file into startup.cvs (akin to ffmpeg.cvs) then GitHub can render it as such which is more readable. :)

[yamt]

do you mean csv?

[Robbepop]

yes sorry

[Robbepop]

At least for me GitHub beautifully renders your ffmpeg.csv.