This repository shall contain various experiments, API tests, benchmarks etc. Anything worth keeping should be copied to the main codebase.
Sometimes it's not really worth it to setup an entire app just to test performance of a specific API.
- StructuredClone vs. Transfer to a Worker
- sending ArrayBuffers to a Worker with
postMessage, either let the runtime work its magic or manually clone and transfer it - clone (
buf.slice()) and transfer is fastest but by surprisingly little margin (1600 vs. 1300 op/s, 20%) - contents of the buffer don't seem to make much difference (rand vs. zero)
- cloning and copying is stupid ...
- sending ArrayBuffers to a Worker with
- Many Messages on MessageChannels
- creating 100
MessageChannels and sending 10000 messages - either all on a single channel or spread across all channels evenly
- no matter how the
chans/msgsvariables are set, the even spread is consistently faster (8.03 op/s ±3.50% vs. 4.75 op/s ±2.73%) - the bench does not create workers, so this is simply the efficiency of message handling on a single thread
- creating 100
- Many Messages to Workers
- interestingly, the same does not hold for multiple Workers ..
- sending all messages to a single Worker is faster (7.69 op/s ±6.99%) than spreading it across many (3.72 op/s ±2.99%)
- so, even with an optimized Comlink, I will hit a performance bottleneck here
- although, we're doing between 372.000 and 769.000 requests per second here ...