📖 Docs

Top-K Heavykeeper algorithm for Top-K elephant flows

This is based on the paper HeavyKeeper: An Accurate Algorithm for Finding Top-k Elephant Flows by Junzhi Gong, Tong Yang, Haowei Zhang, and Hao Li, Peking University; Steve Uhlig, Queen Mary, University of London; Shigang Chen, University of Florida; Lorna Uden, Staffordshire University; Xiaoming Li, Peking University

See examples/basic.rs for a complete example, or examples/ip_files.rs for an example of counting top-k IP flows in a file.

Basic usage:

use heavykeeper::TopK;

// create a new TopK with k=10, width=1000, depth=4, decay=0.9
let mut topk: TopK<String> = TopK::new(10, 1000, 4, 0.9);

// add some items
topk.add("example item", 5);
topk.add("another item", 1);

// check the counts
for node in topk.list() {
    println!("{} {}", node.item, node.count);
}

The crate ships three top-K sketches that share the same public API (new / with_seed / with_hasher / builder / add / count / contains / list / merge):

Numbers are from cargo bench --bench topk_vs_bucketed at K=100, width=4096, depth=4, decay=0.9 on u64 keys. Recall is from tests/accuracy_compare.rs (paper-style heavy-hitter test, φ = 0.0005, 1 M Zipf samples).

TopK is the canonical implementation from the paper, with its accuracy bounds. BucketedTopK and CuckooTopK are derived variants — they don't carry the paper's row-independence accuracy bounds, but the empirical accuracy on Zipf streams is competitive and often better.

Pick by workload:

• TopK — when you want the published algorithm and its bounds.
• BucketedTopK — best general-purpose insert throughput; closest to TopK's cost model with a single bucket per key.
• CuckooTopK — best accuracy and throughput on heavy-hitter-skewed traffic (the elephant-flow use case). Each bucket has a single lobby cell with probabilistic decay plus depth non-decaying heavy slots; promoted items live in one of two cuckoo candidate buckets and are re-homed on collision via a kick chain (bound configurable via CuckooBuilder::max_kicks, default 8).

All three support seedable construction, custom hashers, and merge between compatible instances. Errors are returned via BuilderError/MergeError enums; the infallible constructors (new, with_seed, with_hasher) trust the caller. Bucket indexing uses an AND-mask fast-path when width is a power of two; pick power-of-two widths in production for the best per-add cost.

examples/ip_files.rs runs all three sketches over a CAIDA-style trace (27.5 M packets, 1.03 M distinct 13-byte flow keys = src IP : src port → dst IP : dst port + protocol). Same K=1000, decay=0.95, equal cell budgets across variants:

A word count program that demonstrates the HeavyKeeper algorithm can be found at examples/word_count.rs.

cargo build --example word_count --release
target/release/examples/word_count -k 10 -w 8192 -d 2 -y 0.95 -f data/war_and_peace.txt

cargo run --example basic --release

cargo run --example ip_files --release

cargo bench

hyperfine 'target/release/examples/word_count -k 10 -w 8192 -d 2 -y 0.95 -f data/war_and_peace.txt'

For information about test data format and how to obtain or generate test data, please see data/README.md.

This project is dual licensed under the Apache/MIT license.