WebAssembly Runtime Performance

Note: A major benefit of WebAssembly is startup time as parsing of wasm is more efficient than parsing JS. The work here doesn't take that into account. It only looks at performance of actually executing the code after everything is loaded.

You can run the following benchmarks at:

• https://azeemba.github.io/wasm-runtime-performance/output/data_O3.html (optimized C++)
• https://azeemba.github.io/wasm-runtime-performance/output/data.html (unoptimized)

Floyd Warshall

Thing being compared:

• WASM based:
  • boost::Graph's implementation of Floyd Warshall
• graphlib.js

The results here are very interesting. Here is the output when web assembly is generated with no optimization:

FloydWarshall in JS with V: 100 E: 500 took in ms: 62.16005000000006
FloydWarshall in WASM with V: 100 E: 500 took in ms: 1235.4100999999998

Here is the output when web assembly is generated at -O3 level:

FloydWarshall in JS with V: 100 E: 500 took in ms: 61.1135
FloydWarshall in WASM with V: 100 E: 500 took in ms: 10.156900000000315

The performance of wasm improved by 3 orders of magnitude! It went from being much worse than the JS to being much better!

Heap/PriorityQueue Performance:

Things being compared:

• WASM based:
  • std::priority_queue with individual actions
  • std::priority_queue with bulk actions
• FastPriorityQueue.js
• Heap.js

Here is a representative output at -O3 optimized level (smaller number is good):

In main
WASM:
On average, inserts of 300000 elems took in ms:  58.43675000000002
On average, peek+pop of 300000 elems took in ms:  39.92820000000006

WASM Bulk Operations: 
On average, inserts of 300000 elems took in ms:  51.98654999999981
On average, peek+pop of 300000 elems took in ms:  33.065599999999684

FastPriorityQueue.js: 
On average, inserts of 300000 elems took in ms:  9.860999999999585
On average, peek+pop of 300000 elems took in ms:  27.373649999999945

Heap.js: 
On average, inserts of 300000 elems took in ms:  13.4233500000001
On average, peek+pop of 300000 elems took in ms:  32.30655000000024

So it looks like web-assembly definitely does have a performance hit. Though surprisingly the behavior is a lot less significant when pop-ing the elements.

Maintainence:

Requires boost and emsdk The wasm, html, js can be generated via ./generate_output.sh

To do:

Compare performance of other data structures and algorithms to get a better holistic view