Progress report in Pen programming language

June 19th, 2022

@raviqqe

Agenda

• Progress report
  • Relaxed atomic operations in reference counting
• Next plans

Progress report

Relaxed atomic operations in reference counting

• Reference counting can use relaxed atomic or non-atomic operations sometimes.
  • For only references never shared by multiple threads
• Part of the Perceus reference counting algorithm

Benchmark

• Conway's game of life
• Size: 20 x 40
• Iterations: 100

> hyperfine -w 3 ~/app-old ~/app-new
Benchmark 1: /home/raviqqe/app-old
  Time (mean ± σ):      1.712 s ±  0.011 s    [User: 1.658 s, System: 0.012 s]
  Range (min … max):    1.694 s …  1.727 s    10 runs

Benchmark 2: /home/raviqqe/app-new
  Time (mean ± σ):      1.175 s ±  0.021 s    [User: 1.124 s, System: 0.008 s]
  Range (min … max):    1.152 s …  1.229 s    10 runs

Summary
  '/home/raviqqe/app-new' ran
    1.46 ± 0.03 times faster than '/home/raviqqe/app-old'

Record update benchmark

• The previous result had several performance bugs!
  • Inefficient map literal compilation
  • Non-unique references

Configuration

• Hash map initialization with many entries
• A number of entries: 100,000
• Key type: 64-bit floating point number

Results

• Pen is 6 ~ 7 times slower than Rust currently...

Pen

> hyperfine -w 3 ./app
Benchmark 1: ./app
  Time (mean ± σ):     274.0 ms ±   2.7 ms    [User: 206.7 ms, System: 16.8 ms]
  Range (min … max):   269.6 ms … 279.3 ms    10 runs

im-rs in Rust

• HashMap::insert(&mut self, k: K, v: V) -> Option<V>

test hashmap_insert_mut_100000 ... bench:  34,869,571 ns/iter (+/- 4,337,627)

Next plans

• Reference counting optimization
  • Unboxing small records #671
  • Other basic optimizations
• Proper C calling convention in FFI #444
  • Compiling to MLIR?

Summary

• Progress
  • Relaxed atomic operations in reference counting
• Next plans
  • Record unboxing

Appendix

Record update benchmark

No uniqueness check

> for _ in $(seq 5); do time ./app; done
./app  7.98s user 0.26s system 99% cpu 8.302 total
./app  7.70s user 0.24s system 99% cpu 7.991 total
./app  7.71s user 0.29s system 99% cpu 8.052 total
./app  7.76s user 0.31s system 99% cpu 8.117 total
./app  8.10s user 0.26s system 99% cpu 8.423 total

Acquire ordering

> for _ in $(seq 5); do time ./app; done
./app  7.68s user 0.28s system 99% cpu 8.019 total
./app  7.57s user 0.32s system 99% cpu 7.950 total
./app  7.63s user 0.22s system 99% cpu 7.905 total
./app  7.58s user 0.26s system 99% cpu 7.899 total
./app  7.59s user 0.27s system 99% cpu 7.933 total

Relaxed (buggy) ordering

> for _ in $(seq 5); do time ./app; done
./app  7.46s user 0.30s system 99% cpu 7.817 total
./app  7.41s user 0.24s system 99% cpu 7.703 total
./app  7.51s user 0.26s system 99% cpu 7.823 total
./app  7.47s user 0.26s system 99% cpu 7.775 total
./app  7.42s user 0.26s system 99% cpu 7.734 total

Relaxed (correct) ordering

> for _ in $(seq 5); do time ./app; done
./app  7.60s user 0.26s system 99% cpu 7.930 total
./app  7.50s user 0.29s system 99% cpu 7.860 total
./app  7.60s user 0.27s system 99% cpu 7.940 total
./app  7.56s user 0.25s system 99% cpu 7.865 total
./app  7.55s user 0.27s system 99% cpu 7.878 total

Game of life benchmark

• Relaxed atomic operations for thunks

Before:

> hyperfine ./app
Benchmark 1: ./app
  Time (mean ± σ):     11.637 s ±  0.191 s    [User: 11.588 s, System: 0.094 s]
  Range (min … max):   11.311 s … 11.992 s    10 runs

After:

> hyperfine ./app
Benchmark 1: ./app
  Time (mean ± σ):     11.891 s ±  0.146 s    [User: 11.822 s, System: 0.109 s]
  Range (min … max):   11.614 s … 12.097 s    10 runs