Traverse without recursion

[RobinTail]

Avoiding recursion should decrease memory consumption by replacing the utilization of call stack to a stack of entities to process

https://www.nearform.com/insights/tracking-memory-allocation-node-js/

   ✓ Experiment for routing walker (2) 1247ms
     name             hz     min     max    mean     p75     p99    p995    p999     rme  samples
   · current    3,691.20  0.2053  2.7257  0.2709  0.2885  0.3817  0.4519  2.1826  ±1.49%     1846
   · featured  50,253.26  0.0123  5.8032  0.0199  0.0165  0.0423  0.1116  2.7895  ±7.10%    25127   fastest

Turned out that the performance is mostly affected by assert(..., new Error()) pattern, partially introduced by #1371

[coveralls-official]

coverage: 100.0%. remained the same
when pulling 36608df on traverse-no-recursion
into b8f0906 on make-v21.

[RobinTail]

After fixing in #2169

   ✓ Experiment for routing walker (2) 1269ms
     name             hz     min      max    mean     p75     p99    p995    p999     rme  samples
   · old       46,289.59  0.0114   5.7957  0.0216  0.0179  0.0497  0.1211  2.8412  ±6.92%    23145
   · featured  48,947.95  0.0132  10.0317  0.0204  0.0168  0.0273  0.0357  3.1846  ±8.16%    24474   fastest

[RobinTail]

Evaluation using --trace-gc in Node 22 with a help of ChatGPT

tsx runner involved in both cases

Key Observations

1. Scavenge Performance:

   • Old Logs: Scavenge cycles often exceed 1 ms (e.g., 1.19 ms, 1.33 ms) with notable occurrences of allocation failures leading to fallback GC strategies.
   • New Logs: Scavenge cycles are generally faster (mostly <1 ms, e.g., 0.59 ms, 0.78 ms). Allocation failures still occur but are less frequent and seem to have reduced overhead.

2. Mark-Compact Operations:

   • Old Logs: Mark-Compact cycles appear more prolonged, with finalization phases requiring more time (e.g., 3.41 ms, 3.01 ms) and frequent triggering in response to allocation failures.
   • New Logs: Mark-Compact cycles are optimized in terms of latency (e.g., 2.08 ms, 3.01 ms). The new implementation shows slightly fewer allocation failures requiring Mark-Compact operations.

3. Memory Usage:

   • Old Logs: The system uses more memory (e.g., up to 67.1 MB post-GC in old space).
   • New Logs: Memory usage appears better managed, with memory peaking slightly lower (e.g., 66.8 MB post-GC in old space).

4. Overall Efficiency:

   • Old Logs: The system experiences higher cumulative latency and allocation failures, implying more frequent reliance on fallback strategies like Mark-Compact GC.
   • New Logs: Improved handling of allocation failures and faster scavenges suggest better memory management and reduced GC overhead.

---

Conclusions

• Improvement in Scavenge Cycles: The new system shows significant improvement in Scavenge GC cycles, leading to faster recovery of young-generation memory.
• Reduced Allocation Failures: Allocation failures still occur but are better handled, resulting in smoother transitions and fewer interruptions.
• Optimized Mark-Compact GC: The Mark-Compact process, while still necessary, is more streamlined and efficient in the new system.

[RobinTail]

Summary of Improvements

Memory Usage Reduction: ~17.14%
GC Time Reduction: ~19.16%

[RobinTail]

ok