Rui Codex entered all four tracks of Hackathon Raptors' Speed Demon 2026 in Rust, standard library only. Code and per-track logs with board run IDs: https://github.com/ruicodex-dev/speed-demon-2026 (MIT).
The one lesson. Our largest step after the first Rust port came from deliberately breaking the program. On Track C (a CSV join) we removed name interning, which makes the output wrong, just to time it: user CPU fell from 2.1 s to 0.55 s. Three quarters of the work was name lookups in twelve per-thread tables. Records now append to per-partition buffers, and each partition owner interns a table that stays in cache. K rose 1.41x (9.40 -> 13.30): 0.165 -> 0.109 s, 139 -> 69 MB. One wrong-on-purpose run showed the ceiling before we built anything.
Setup. The scoring box is a Core i5-13500 pinned to six P-cores (12 threads), turbo off at 2.5 GHz, 8 GB, 1.25 MB of L2 per core. Our laptop (i9-13900H, WSL2, turbo on, 2 MB L2) was shared with other work, so it only ever compared two binaries; the live board decided. K = (T_ref/T)^0.5 x (M_ref/M)^0.2 x (S_ref/S)^0.15 x (D_ref/D)^0.15; points = 50 x log10(K).
The journey (selected steps; every row is a board run in the track's PERF.md).
| Track | Hypothesis | Before | After | Kept? |
|---|---|---|---|---|
| A logs | One per-line array, records rewritten in place | 0.088 s, 74.3 MB | 0.072 s, 42.5 MB | yes |
| A | ~25 % of time in libc memcpy/memcmp: masked 8-byte loads | 0.0613 s | 0.0555 s | yes |
| B search | Phrases matched by Aho-Corasick during the scan | 1.63 s | 1.14 s | yes |
| B | Frequent terms stored as bitmaps | 0.78 s, 300 MB | 0.61 s, 187 MB | yes |
| C join | Sequential appends + cache-sized tables | 0.165 s, 139 MB | 0.109 s, 69 MB | yes |
| C | Slot index as name id (laptop: 1.13x faster) | 0.1016 s, 39.9 MB | 0.1081 s, 44.7 MB | no |
| D paths | Targets with no in-edge answered without a search | 1.43 s | 0.96 s | yes |
| D | Nodes renumbered in BFS order | 0.96 s | 0.58 s | yes |
| D | One u32 per edge, rare heavy weights in a map | 118.0 MB | 81.1 MB | yes |
What did not work.
A laptop win the board refused. C's slot-as-id change removed one of three cache misses per lookup and was 1.13x faster locally. On the board it was slower and bigger: the bitsets grew 2.6x, and the scoring CPU has 1.25 MB of L2 against our 2 MB. Reverted the next commit.
Fewer nodes, worse score. Bidirectional Dijkstra on D settles 42 % fewer nodes. The first version was 2.3x slower; perf found three causes in turn (arrays filled instead of calloc'd, u32 overflow on the spec's near-2^32 edges, one side sweeping the whole graph). Fixed, it was 1.14x faster, but peak memory went 146 -> 258 MiB: 1.14^0.5 x (146/258)^0.2 = 0.95. Kept out.
A measurement we misread. Chunk-local bitmaps on B, replacing ~25M atomic fetch_or calls, showed no gain in a wall-clock A/B (1.105 +/- 0.264 s against 1.021 +/- 0.186 s). perf later put that lock or at 8.6 %. Interleaving the two binaries run by run and comparing median CPU: 18.2 -> 16.7 s. Sequential A/Bs of one pair had read 0.83 vs 1.16 s, then 1.69 vs 1.14 s, so every later A/B was interleaved.
One call, opposite effects. The board read about 25 MB more than our laptop on C. Smaller thread stacks and a single malloc arena changed nothing; setting transparent huge pages to always locally reproduced it. prctl(PR_SET_THP_DISABLE) took C from 41.0 to 16.8 MB without costing time, but made B (0.48 -> 0.52 s) and D (0.30 -> 0.34 s) slower. Only C keeps it.
Tune for 12 threads, not yours. At the board's 12 threads a D build bucket was ~63.5 KB, just under glibc's 64 KB mmap threshold, so freeing it returned nothing (75 KB at the laptop's 10 threads). A threshold derived from the bucket size took board memory from 79-81 to 75 MB.
Correctness. The gate is byte-for-byte output, five runs out of five. Every pushed change was also checked on seeds 2 and 3, seed 2 at full size, and tiny seeds for A and C. Shortcuts carry exact fallbacks: D's u32 distances repeat in u64 on overflow; C's name table restarts four times larger if it fills; B's frequency sample picks a representation, never an answer. C's chunking relies on a spec guarantee (CR only in CRLF terminators), not on the public file. The final commit passed every gate on all four tracks.
K, not just time. Growing the source by f pays only if the speed-up beats f^0.3: +50 % lines needs 1.13x. No track has a dependency (D = 1). Memory was often the cheapest win: A maps its input instead of copying it, and frees its records before writing its 10 MB output, whose page cache is charged to the run.
Results. Reference -> final commit: A 32.05 s -> 0.0552 s, B 56.35 -> 0.490 s, C 14.44 -> 0.0750 s, D 72.69 -> 0.235 s; memory 228 -> 25.5, 2,200 -> 122.4, 367.4 -> 12.4 and 892.9 -> 76.1 MB. On the frozen public board we were mid-table: A 71st of 134, B 66th of 104, C 75th of 111, D 40th of 113. Final scores (holdout seed plus judges) come on 13 November. Noise is real: the same A code ran 0.0490 s and, as our final commit, 0.0552 s.
With 24 more hours. D: the query loop is 69 % of CPU, mostly misses on neighbour distances (16 %) and offsets (~8 %); it needs a search order that does not double memory. B, our weakest track: profile-guided optimisation was allowed and is not in our log. A: the final 10 MB write is serial (9-18 ms locally).
How it was built. With Claude Code, Anthropic's coding agent: it wrote the Rust, ran perf and the interleaved A/Bs, and drafted this post. Every change was checked against the expected output hashes, measured before it was kept, and reviewed by Rui Codex; the reverted ones are in the logs too.
Sources cited in this article
- https://github.com/ruicodex-dev/speed-demon-2026 (github.com)