Performance: Actor/strand spawn is 20-2500x slower than alternatives #307

New issue

Closed

opened 2026-01-25 23:13:12 +00:00 by navicore · 1 comment

navicore commented

2026-01-25 23:13:12 +00:00

(Migrated from github.com)

Current State

Skynet benchmark: spawn 100k actors in 10-ary tree, collect sum

Language	Time	vs Seq
Seq	5,165ms	1x
Python (asyncio)	266ms	20x faster
Go	19ms	270x faster
Rust	2ms	2,500x faster

Root Causes

Strand allocation: Each strand has significant setup cost
Channel per strand: Every actor creates a channel for results
Thread pool scheduling: Work distribution overhead
Deep recursion: 10-ary tree creates stack pressure

Potential Approaches

Near-term

Lighter strand creation: Reduce per-strand allocation
Channel pooling: Reuse channels for short-lived communications
Work stealing: Better load balancing across threads

Long-term

Stackless coroutines: Lighter than full strand state machines
Strand pooling: Reuse strand slots instead of allocating new
Spawn fusion: Batch multiple spawns into single scheduling operation

Benchmark Code

: skynet ( Channel Int Int -- )
  dup 1 i.= if
    drop swap chan.send drop
  else
    chan.make
    over 10 i.divide drop
    # Spawn 10 children
    3 pick 10 i.* 0 i.+ 2 pick swap 2 pick [ skynet ] strand.spawn drop drop drop drop
    3 pick 10 i.* 1 i.+ 2 pick swap 2 pick [ skynet ] strand.spawn drop drop drop drop
    # ... (10 spawns total)
    
    # Collect 10 results
    dup chan.receive drop
    over chan.receive drop i.+
    # ... (10 receives total)
  then
;

Success Criteria

Spawn 100k actors in < 500ms (10x improvement)
Within 50x of Go for tree-structured spawns

## Current State Skynet benchmark: spawn 100k actors in 10-ary tree, collect sum | Language | Time | vs Seq | |----------|------|--------| | Seq | 5,165ms | 1x | | Python (asyncio) | 266ms | 20x faster | | Go | 19ms | 270x faster | | Rust | 2ms | 2,500x faster | ## Root Causes 1. **Strand allocation**: Each strand has significant setup cost 2. **Channel per strand**: Every actor creates a channel for results 3. **Thread pool scheduling**: Work distribution overhead 4. **Deep recursion**: 10-ary tree creates stack pressure ## Potential Approaches ### Near-term - **Lighter strand creation**: Reduce per-strand allocation - **Channel pooling**: Reuse channels for short-lived communications - **Work stealing**: Better load balancing across threads ### Long-term - **Stackless coroutines**: Lighter than full strand state machines - **Strand pooling**: Reuse strand slots instead of allocating new - **Spawn fusion**: Batch multiple spawns into single scheduling operation ## Benchmark Code ```seq : skynet ( Channel Int Int -- ) dup 1 i.= if drop swap chan.send drop else chan.make over 10 i.divide drop # Spawn 10 children 3 pick 10 i.* 0 i.+ 2 pick swap 2 pick [ skynet ] strand.spawn drop drop drop drop 3 pick 10 i.* 1 i.+ 2 pick swap 2 pick [ skynet ] strand.spawn drop drop drop drop # ... (10 spawns total) # Collect 10 results dup chan.receive drop over chan.receive drop i.+ # ... (10 receives total) then ; ``` ## Success Criteria - Spawn 100k actors in < 500ms (10x improvement) - Within 50x of Go for tree-structured spawns

navicore commented

2026-03-23 21:50:15 +00:00

(Migrated from github.com)

https://github.com/navicore/patch-seq/pull/367