Use custom d-ary heap implementation #7017

SiarheiFedartsou · 2024-08-19T19:04:01Z

Context

Benchmark highlights

CH Map Matching +16% ops/sec
MLD Map matching +30% ops/sec
CH Routing +3% ops/sec
MLD Routing +18% ops/sec
CH Table +7% ops/sec
MLD Table +16% ops/sec

osrm_contract run time on berlin-latest.osm.pbf -24%
osrm_customize run time on berlin-latest.osm.pbf -25%

Description

It turns out that with mutability enabled boost::heap::d_ary_heap becomes extremely ineffective. I wrote such benchmark which compares boost::heap::mutable_<true> vs boost::heap::mutable_<false>:

    const int NUM_ELEMENTS = 10000000;
    std::vector<uint16_t> data;
    data.reserve(NUM_ELEMENTS);

    std::mt19937 gen(42);
    std::uniform_int_distribution<> dis(1, 1000000);

    for (int i = 0; i < NUM_ELEMENTS; ++i) {
        data.push_back(dis(gen));
    }

    std::vector<int> resultBoostDAryHeapMutable;
    resultBoostDAryHeapMutable.reserve(NUM_ELEMENTS);
    auto start = std::chrono::high_resolution_clock::now();
    boost::heap::d_ary_heap<uint16_t, boost::heap::arity<4>, boost::heap::mutable_<true>, boost::heap::compare<std::greater<uint16_t>>> boostDAryHeapMutable;
    for (const auto& val : data) {
        boostDAryHeapMutable.push(val);
    }
    while (!boostDAryHeapMutable.empty()) {
        resultBoostDAryHeapMutable.push_back(boostDAryHeapMutable.top());
        boostDAryHeapMutable.pop();
    }
    auto end = std::chrono::high_resolution_clock::now();
    auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count();
    std::cout << "Boost d_ary_heap mutable took " << duration << " ms" << std::endl;

    std::vector<int> resultBoostDAryHeap;
    resultBoostDAryHeap.reserve(NUM_ELEMENTS);
    start = std::chrono::high_resolution_clock::now();
    boost::heap::d_ary_heap<uint16_t, boost::heap::arity<4>, boost::heap::compare<std::greater<uint16_t>>> boostDAryHeap;
    for (const auto& val : data) {
        boostDAryHeap.push(val);
    }
    while (!boostDAryHeap.empty()) {
        resultBoostDAryHeap.push_back(boostDAryHeap.top());
        boostDAryHeap.pop();
    }
    end = std::chrono::high_resolution_clock::now();
    duration = std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count();
    std::cout << "Boost d_ary_heap took " << duration << " ms" << std::endl;
}

And it turns out that mutable implementation(i.e. where we can decrease priorities) is 10-12x(!) slower than non-mutable one:

Boost d_ary_heap mutable took 11142 ms
Boost d_ary_heap took 901 ms

Implementation of boost::heap::d_ary_heap is quite difficult to read, but as I understood in order to achieve mutability it uses std::list under the hood what obviously makes it quite slow(at least due to additional allocations and non cache-friendliness).
So I decided to make our own implementation of 4-ary heap which handles mutability in a different way: instead of having stable iterators we use indexes in heap's underlying vector with a bit of additional logic to update stored indexes when elements are reordered within the heap.

Benchmark Results

Benchmark	Base	PR
alias	aliased u32: 11076 plain u32: 10972.9 aliased double: 15129.3 plain double: 15113.2	aliased u32: 11009.8 plain u32: 10996.4 aliased double: 15039.1 plain double: 15025.3
e2e_match_ch	Ops: 25.55 ± 0.01 ops/s. Best: 25.57 ops/s Total: 5126.51ms ± 2.43ms. Best: 5122.78ms Min time: 3.05ms ± 0.05ms Mean time: 39.13ms ± 0.02ms Median time: 27.37ms ± 0.08ms 95th percentile: 132.54ms ± 0.35ms 99th percentile: 158.85ms ± 0.46ms Max time: 179.68ms ± 0.75ms	Ops: 30.69 ± 0.03 ops/s. Best: 30.75 ops/s Total: 4268.10ms ± 5.06ms. Best: 4259.72ms Min time: 2.93ms ± 0.05ms Mean time: 32.58ms ± 0.04ms Median time: 22.89ms ± 0.14ms 95th percentile: 106.88ms ± 0.12ms 99th percentile: 128.44ms ± 0.57ms Max time: 141.70ms ± 0.89ms
e2e_match_mld	Ops: 43.57 ± 0.02 ops/s. Best: 43.60 ops/s Total: 3006.32ms ± 1.34ms. Best: 3004.43ms Min time: 2.58ms ± 0.05ms Mean time: 22.95ms ± 0.01ms Median time: 12.06ms ± 0.12ms 95th percentile: 75.20ms ± 0.12ms 99th percentile: 86.91ms ± 0.14ms Max time: 100.71ms ± 0.30ms	Ops: 59.95 ± 0.05 ops/s. Best: 60.03 ops/s Total: 2184.99ms ± 1.71ms. Best: 2182.09ms Min time: 2.24ms ± 0.02ms Mean time: 16.68ms ± 0.01ms Median time: 9.72ms ± 0.04ms 95th percentile: 52.05ms ± 0.10ms 99th percentile: 60.11ms ± 0.16ms Max time: 69.53ms ± 0.39ms
e2e_nearest_ch	Ops: 635.40 ± 2.55 ops/s. Best: 640.86 ops/s Total: 1573.68ms ± 6.29ms. Best: 1560.39ms Min time: 1.28ms ± 0.01ms Mean time: 1.57ms ± 0.01ms Median time: 1.54ms ± 0.00ms 95th percentile: 1.96ms ± 0.01ms 99th percentile: 2.05ms ± 0.01ms Max time: 9.38ms ± 7.45ms	Ops: 636.59 ± 3.36 ops/s. Best: 642.66 ops/s Total: 1570.80ms ± 8.30ms. Best: 1556.04ms Min time: 1.28ms ± 0.01ms Mean time: 1.57ms ± 0.01ms Median time: 1.53ms ± 0.01ms 95th percentile: 1.96ms ± 0.01ms 99th percentile: 2.04ms ± 0.01ms Max time: 9.32ms ± 7.40ms
e2e_nearest_mld	Ops: 635.02 ± 4.27 ops/s. Best: 640.52 ops/s Total: 1574.46ms ± 10.06ms. Best: 1561.23ms Min time: 1.29ms ± 0.01ms Mean time: 1.57ms ± 0.01ms Median time: 1.54ms ± 0.01ms 95th percentile: 1.97ms ± 0.01ms 99th percentile: 2.05ms ± 0.01ms Max time: 9.39ms ± 7.43ms	Ops: 636.33 ± 2.53 ops/s. Best: 639.88 ops/s Total: 1571.45ms ± 6.62ms. Best: 1562.79ms Min time: 1.28ms ± 0.01ms Mean time: 1.57ms ± 0.01ms Median time: 1.54ms ± 0.00ms 95th percentile: 1.96ms ± 0.00ms 99th percentile: 2.05ms ± 0.01ms Max time: 9.28ms ± 7.37ms
e2e_route_ch	Ops: 217.14 ± 0.62 ops/s. Best: 218.12 ops/s Total: 4605.59ms ± 13.80ms. Best: 4584.62ms Min time: 1.83ms ± 0.03ms Mean time: 4.61ms ± 0.01ms Median time: 4.70ms ± 0.01ms 95th percentile: 6.06ms ± 0.03ms 99th percentile: 6.60ms ± 0.04ms Max time: 13.34ms ± 6.42ms	Ops: 218.93 ± 0.47 ops/s. Best: 219.97 ops/s Total: 4567.70ms ± 10.21ms. Best: 4546.15ms Min time: 1.84ms ± 0.05ms Mean time: 4.57ms ± 0.01ms Median time: 4.66ms ± 0.01ms 95th percentile: 6.01ms ± 0.03ms 99th percentile: 6.55ms ± 0.03ms Max time: 13.24ms ± 6.37ms
e2e_route_mld	Ops: 178.18 ± 0.27 ops/s. Best: 178.71 ops/s Total: 5612.39ms ± 8.61ms. Best: 5595.71ms Min time: 1.78ms ± 0.04ms Mean time: 5.61ms ± 0.01ms Median time: 5.73ms ± 0.02ms 95th percentile: 7.61ms ± 0.03ms 99th percentile: 8.15ms ± 0.04ms Max time: 14.77ms ± 5.89ms	Ops: 186.43 ± 0.32 ops/s. Best: 187.08 ops/s Total: 5363.86ms ± 9.38ms. Best: 5345.20ms Min time: 1.78ms ± 0.03ms Mean time: 5.36ms ± 0.01ms Median time: 5.47ms ± 0.02ms 95th percentile: 7.25ms ± 0.01ms 99th percentile: 7.82ms ± 0.06ms Max time: 14.42ms ± 5.98ms
e2e_table_ch	Ops: 218.31 ± 0.48 ops/s. Best: 219.32 ops/s Total: 4580.74ms ± 10.97ms. Best: 4559.55ms Min time: 2.46ms ± 0.04ms Mean time: 4.58ms ± 0.01ms Median time: 4.58ms ± 0.01ms 95th percentile: 6.23ms ± 0.02ms 99th percentile: 6.58ms ± 0.04ms Max time: 13.70ms ± 7.04ms	Ops: 226.94 ± 0.38 ops/s. Best: 227.56 ops/s Total: 4406.36ms ± 7.58ms. Best: 4394.42ms Min time: 2.40ms ± 0.03ms Mean time: 4.41ms ± 0.01ms Median time: 4.40ms ± 0.02ms 95th percentile: 5.96ms ± 0.02ms 99th percentile: 6.29ms ± 0.03ms Max time: 13.45ms ± 7.07ms
e2e_table_mld	Ops: 69.13 ± 0.03 ops/s. Best: 69.16 ops/s Total: 14466.36ms ± 6.86ms. Best: 14460.18ms Min time: 6.01ms ± 0.04ms Mean time: 14.47ms ± 0.01ms Median time: 14.39ms ± 0.01ms 95th percentile: 22.07ms ± 0.04ms 99th percentile: 23.22ms ± 0.08ms Max time: 29.71ms ± 5.67ms	Ops: 78.73 ± 0.02 ops/s. Best: 78.76 ops/s Total: 12701.45ms ± 4.01ms. Best: 12696.81ms Min time: 5.38ms ± 0.02ms Mean time: 12.70ms ± 0.00ms Median time: 12.65ms ± 0.01ms 95th percentile: 19.31ms ± 0.06ms 99th percentile: 20.29ms ± 0.06ms Max time: 26.86ms ± 5.89ms
e2e_trip_ch	Ops: 62.95 ± 0.03 ops/s. Best: 63.00 ops/s Total: 15885.05ms ± 8.02ms. Best: 15874.26ms Min time: 2.43ms ± 0.22ms Mean time: 15.88ms ± 0.01ms Median time: 15.10ms ± 0.03ms 95th percentile: 28.09ms ± 0.03ms 99th percentile: 30.17ms ± 0.09ms Max time: 32.49ms ± 1.45ms	Ops: 64.23 ± 0.04 ops/s. Best: 64.29 ops/s Total: 15568.75ms ± 9.29ms. Best: 15553.72ms Min time: 2.38ms ± 0.17ms Mean time: 15.57ms ± 0.01ms Median time: 14.78ms ± 0.03ms 95th percentile: 27.65ms ± 0.05ms 99th percentile: 29.60ms ± 0.09ms Max time: 32.03ms ± 1.63ms
e2e_trip_mld	Ops: 36.66 ± 0.01 ops/s. Best: 36.68 ops/s Total: 27281.21ms ± 8.68ms. Best: 27263.72ms Min time: 2.42ms ± 0.17ms Mean time: 27.28ms ± 0.01ms Median time: 26.38ms ± 0.05ms 95th percentile: 44.34ms ± 0.05ms 99th percentile: 47.07ms ± 0.12ms Max time: 49.41ms ± 0.11ms	Ops: 40.05 ± 0.02 ops/s. Best: 40.08 ops/s Total: 24966.90ms ± 14.12ms. Best: 24951.03ms Min time: 2.45ms ± 0.19ms Mean time: 24.97ms ± 0.01ms Median time: 24.13ms ± 0.07ms 95th percentile: 41.08ms ± 0.04ms 99th percentile: 43.63ms ± 0.11ms Max time: 45.98ms ± 0.42ms
json-render	String: 8.89322ms Stringstream: 14.6254ms Vector: 9.48771ms	String: 9.15423ms Stringstream: 14.6142ms Vector: 9.53894ms
match_ch	Default radius: 7.06186ms/req at 82 coordinate 0.0861202ms/coordinate Radius 10m: 25.0142ms/req at 82 coordinate 0.305052ms/coordinate	Default radius: 6.28653ms/req at 82 coordinate 0.076665ms/coordinate Radius 10m: 22.18ms/req at 82 coordinate 0.270488ms/coordinate
match_mld	Default radius: 4.35379ms/req at 82 coordinate 0.053095ms/coordinate Radius 10m: 16.2432ms/req at 82 coordinate 0.198088ms/coordinate	Default radius: 3.11549ms/req at 82 coordinate 0.0379938ms/coordinate Radius 10m: 10.7051ms/req at 82 coordinate 0.13055ms/coordinate
node_match_ch	Ops: 159.3 ± 0.5 ops/s. Best: 160.0 ops/s	Ops: 186.1 ± 1.3 ops/s. Best: 187.3 ops/s
node_match_mld	Ops: 219.5 ± 2.2 ops/s. Best: 221.6 ops/s	Ops: 286.7 ± 1.7 ops/s. Best: 290.0 ops/s
node_nearest_ch	Ops: 9989.3 ± 821.8 ops/s. Best: 11065.2 ops/s	Ops: 9393.8 ± 443.1 ops/s. Best: 10209.7 ops/s
node_nearest_mld	Ops: 9286.5 ± 369.8 ops/s. Best: 9970.7 ops/s	Ops: 10275.8 ± 858.9 ops/s. Best: 12036.0 ops/s
node_route_ch	Ops: 1002.6 ± 25.0 ops/s. Best: 1039.0 ops/s	Ops: 1031.5 ± 23.5 ops/s. Best: 1057.3 ops/s
node_route_mld	Ops: 506.0 ± 6.3 ops/s. Best: 515.4 ops/s	Ops: 585.5 ± 3.6 ops/s. Best: 590.5 ops/s
node_table_ch	Ops: 186.4 ± 1.9 ops/s. Best: 188.9 ops/s	Ops: 197.6 ± 1.9 ops/s. Best: 200.9 ops/s
node_table_mld	Ops: 37.8 ± 0.1 ops/s. Best: 37.9 ops/s	Ops: 44.0 ± 0.1 ops/s. Best: 44.1 ops/s
node_trip_ch	Ops: 185.9 ± 1.0 ops/s. Best: 187.0 ops/s	Ops: 193.7 ± 0.4 ops/s. Best: 194.6 ops/s
node_trip_mld	Ops: 61.4 ± 0.2 ops/s. Best: 61.6 ops/s	Ops: 71.8 ± 0.3 ops/s. Best: 72.1 ops/s
osrm_contract	Time: 184.30s Peak RAM: 195.01MB	Time: 141.54s Peak RAM: 195.00MB
osrm_customize	Time: 2.53s Peak RAM: 112.45MB	Time: 1.92s Peak RAM: 112.45MB
osrm_extract	Time: 24.33s Peak RAM: 398.68MB	Time: 24.26s Peak RAM: 399.01MB
osrm_partition	Time: 5.98s Peak RAM: 119.62MB	Time: 5.95s Peak RAM: 119.43MB
packedvector	random write: std::vector 184015 ms util::packed_vector 373595 ms slowdown: 2.03025 random read: std::vector 100548 ms util::packed_vector 190832 ms slowdown: 1.89791	random write: std::vector 184979 ms util::packed_vector 373751 ms slowdown: 2.02051 random read: std::vector 100553 ms util::packed_vector 190728 ms slowdown: 1.89679
random_match_ch	500 matches, default radius ops: 120.21 ± 0.19 ops/s. best: 120.49ops/s. total: 474.18 ± 0.76ms. best: 473.05ms. avg: 8.32 ± 0.01ms min: 0.23 ± 0.01ms max: 41.83 ± 0.17ms p99: 41.83 ± 0.17ms 500 matches, radius=10 ops: 34.48 ± 0.05 ops/s. best: 34.53ops/s. total: 1856.08 ± 2.44ms. best: 1853.21ms. avg: 29.00 ± 0.04ms min: 0.22 ± 0.00ms max: 412.12 ± 1.39ms p99: 412.12 ± 1.39ms 500 matches, radius=20 ops: 8.11 ± 0.01 ops/s. best: 8.12ops/s. total: 8019.57 ± 8.18ms. best: 8009.41ms. avg: 123.38 ± 0.13ms min: 0.49 ± 0.00ms max: 2128.87 ± 3.56ms p99: 2128.87 ± 3.56ms Peak RAM: 55.000MB	500 matches, default radius ops: 146.70 ± 0.27 ops/s. best: 147.08ops/s. total: 388.54 ± 0.73ms. best: 387.54ms. avg: 6.82 ± 0.01ms min: 0.22 ± 0.00ms max: 34.01 ± 0.07ms p99: 34.01 ± 0.07ms 500 matches, radius=10 ops: 42.64 ± 0.04 ops/s. best: 42.68ops/s. total: 1500.93 ± 1.58ms. best: 1499.46ms. avg: 23.45 ± 0.02ms min: 0.22 ± 0.00ms max: 332.39 ± 0.72ms p99: 332.39 ± 0.72ms 500 matches, radius=20 ops: 10.10 ± 0.01 ops/s. best: 10.12ops/s. total: 6436.36 ± 8.71ms. best: 6423.04ms. avg: 99.02 ± 0.13ms min: 0.43 ± 0.00ms max: 1691.74 ± 4.69ms p99: 1691.74 ± 4.69ms Peak RAM: 55.000MB
random_match_mld	500 matches, default radius ops: 205.04 ± 0.49 ops/s. best: 205.56ops/s. total: 278.00 ± 0.67ms. best: 277.29ms. avg: 4.88 ± 0.01ms min: 0.21 ± 0.00ms max: 27.09 ± 0.05ms p99: 27.09 ± 0.05ms 500 matches, radius=10 ops: 73.10 ± 0.06 ops/s. best: 73.17ops/s. total: 875.56 ± 0.73ms. best: 874.64ms. avg: 13.68 ± 0.01ms min: 0.21 ± 0.00ms max: 161.38 ± 0.32ms p99: 161.38 ± 0.32ms 500 matches, radius=20 ops: 15.37 ± 0.01 ops/s. best: 15.38ops/s. total: 4229.16 ± 2.25ms. best: 4225.76ms. avg: 65.06 ± 0.03ms min: 0.29 ± 0.00ms max: 843.11 ± 1.06ms p99: 843.11 ± 1.06ms Peak RAM: 51.500MB	500 matches, default radius ops: 305.47 ± 1.12 ops/s. best: 306.62ops/s. total: 186.60 ± 0.69ms. best: 185.90ms. avg: 3.27 ± 0.01ms min: 0.18 ± 0.00ms max: 18.14 ± 0.05ms p99: 18.14 ± 0.05ms 500 matches, radius=10 ops: 112.95 ± 0.08 ops/s. best: 113.05ops/s. total: 566.63 ± 0.38ms. best: 566.14ms. avg: 8.85 ± 0.01ms min: 0.20 ± 0.00ms max: 100.66 ± 0.10ms p99: 100.66 ± 0.10ms 500 matches, radius=20 ops: 23.63 ± 0.02 ops/s. best: 23.66ops/s. total: 2750.81 ± 1.83ms. best: 2747.78ms. avg: 42.32 ± 0.03ms min: 0.26 ± 0.00ms max: 536.31 ± 0.44ms p99: 536.31 ± 0.44ms Peak RAM: 51.500MB
random_nearest_ch	10000 nearest, number_of_results=1 ops: 21344.30 ± 41.26 ops/s. best: 21377.87ops/s. total: 468.51 ± 0.91ms. best: 467.77ms. avg: 0.05 ± 0.00ms min: 0.02 ± 0.00ms max: 0.16 ± 0.03ms p99: 0.10 ± 0.00ms 10000 nearest, number_of_results=5 ops: 15599.62 ± 3.63 ops/s. best: 15604.45ops/s. total: 641.04 ± 0.15ms. best: 640.84ms. avg: 0.06 ± 0.00ms min: 0.03 ± 0.00ms max: 0.15 ± 0.00ms p99: 0.12 ± 0.00ms 10000 nearest, number_of_results=10 ops: 11959.99 ± 3.89 ops/s. best: 11965.45ops/s. total: 836.12 ± 0.27ms. best: 835.74ms. avg: 0.08 ± 0.00ms min: 0.04 ± 0.00ms max: 0.19 ± 0.00ms p99: 0.15 ± 0.00ms Peak RAM: 34.500MB	10000 nearest, number_of_results=1 ops: 21446.43 ± 39.13 ops/s. best: 21477.18ops/s. total: 466.28 ± 0.85ms. best: 465.61ms. avg: 0.05 ± 0.00ms min: 0.02 ± 0.00ms max: 0.15 ± 0.02ms p99: 0.10 ± 0.00ms 10000 nearest, number_of_results=5 ops: 15616.52 ± 12.28 ops/s. best: 15628.15ops/s. total: 640.35 ± 0.50ms. best: 639.87ms. avg: 0.06 ± 0.00ms min: 0.03 ± 0.00ms max: 0.16 ± 0.02ms p99: 0.13 ± 0.00ms 10000 nearest, number_of_results=10 ops: 11960.74 ± 4.46 ops/s. best: 11965.85ops/s. total: 836.07 ± 0.31ms. best: 835.71ms. avg: 0.08 ± 0.00ms min: 0.04 ± 0.00ms max: 0.19 ± 0.00ms p99: 0.15 ± 0.00ms Peak RAM: 34.500MB
random_nearest_mld	10000 nearest, number_of_results=1 ops: 21347.69 ± 37.14 ops/s. best: 21374.15ops/s. total: 468.44 ± 0.82ms. best: 467.85ms. avg: 0.05 ± 0.00ms min: 0.02 ± 0.00ms max: 0.15 ± 0.02ms p99: 0.10 ± 0.00ms 10000 nearest, number_of_results=5 ops: 15600.41 ± 6.40 ops/s. best: 15606.31ops/s. total: 641.01 ± 0.26ms. best: 640.77ms. avg: 0.06 ± 0.00ms min: 0.03 ± 0.00ms max: 0.15 ± 0.00ms p99: 0.13 ± 0.00ms 10000 nearest, number_of_results=10 ops: 11957.44 ± 6.00 ops/s. best: 11963.98ops/s. total: 836.30 ± 0.43ms. best: 835.84ms. avg: 0.08 ± 0.00ms min: 0.04 ± 0.00ms max: 0.19 ± 0.00ms p99: 0.15 ± 0.00ms Peak RAM: 34.500MB	10000 nearest, number_of_results=1 ops: 21445.62 ± 32.89 ops/s. best: 21472.30ops/s. total: 466.30 ± 0.72ms. best: 465.72ms. avg: 0.05 ± 0.00ms min: 0.02 ± 0.00ms max: 0.15 ± 0.02ms p99: 0.10 ± 0.00ms 10000 nearest, number_of_results=5 ops: 15618.99 ± 6.24 ops/s. best: 15631.75ops/s. total: 640.25 ± 0.26ms. best: 639.72ms. avg: 0.06 ± 0.00ms min: 0.03 ± 0.00ms max: 0.15 ± 0.00ms p99: 0.12 ± 0.00ms 10000 nearest, number_of_results=10 ops: 11947.05 ± 5.39 ops/s. best: 11952.48ops/s. total: 837.03 ± 0.39ms. best: 836.65ms. avg: 0.08 ± 0.00ms min: 0.04 ± 0.00ms max: 0.19 ± 0.00ms p99: 0.15 ± 0.00ms Peak RAM: 34.500MB
random_route_ch	1000 routes, 3 coordinates, no alternatives, overview=full, steps=true ops: 288.66 ± 0.15 ops/s. best: 288.80ops/s. total: 3408.84 ± 1.90ms. best: 3407.15ms. avg: 3.46 ± 0.00ms min: 0.45 ± 0.01ms max: 5.94 ± 0.01ms p99: 5.27 ± 0.01ms 1000 routes, 2 coordinates, 3 alternatives, overview=full, steps=true ops: 324.33 ± 0.05 ops/s. best: 324.42ops/s. total: 3083.29 ± 0.44ms. best: 3082.42ms. avg: 3.08 ± 0.00ms min: 0.12 ± 0.00ms max: 8.35 ± 0.02ms p99: 7.15 ± 0.01ms 1000 routes, 3 coordinates, no alternatives, overview=false, steps=false ops: 597.96 ± 0.06 ops/s. best: 598.05ops/s. total: 1645.59 ± 0.16ms. best: 1645.35ms. avg: 1.67 ± 0.00ms min: 0.35 ± 0.00ms max: 2.68 ± 0.01ms p99: 2.44 ± 0.00ms 1000 routes, 2 coordinates, 3 alternatives, overview=false, steps=false ops: 618.00 ± 0.15 ops/s. best: 618.24ops/s. total: 1618.13 ± 0.40ms. best: 1617.49ms. avg: 1.62 ± 0.00ms min: 0.09 ± 0.00ms max: 5.51 ± 0.01ms p99: 4.01 ± 0.02ms Peak RAM: 83.500MB	1000 routes, 3 coordinates, no alternatives, overview=full, steps=true ops: 294.71 ± 0.09 ops/s. best: 294.81ops/s. total: 3338.92 ± 1.04ms. best: 3337.71ms. avg: 3.39 ± 0.00ms min: 0.43 ± 0.00ms max: 5.80 ± 0.01ms p99: 5.13 ± 0.01ms 1000 routes, 2 coordinates, 3 alternatives, overview=full, steps=true ops: 331.62 ± 0.12 ops/s. best: 331.78ops/s. total: 3015.49 ± 1.06ms. best: 3014.04ms. avg: 3.02 ± 0.00ms min: 0.12 ± 0.00ms max: 8.11 ± 0.01ms p99: 6.97 ± 0.02ms 1000 routes, 3 coordinates, no alternatives, overview=false, steps=false ops: 631.06 ± 0.12 ops/s. best: 631.29ops/s. total: 1559.28 ± 0.31ms. best: 1558.70ms. avg: 1.58 ± 0.00ms min: 0.33 ± 0.00ms max: 2.48 ± 0.01ms p99: 2.35 ± 0.00ms 1000 routes, 2 coordinates, 3 alternatives, overview=false, steps=false ops: 649.07 ± 0.08 ops/s. best: 649.22ops/s. total: 1540.68 ± 0.19ms. best: 1540.32ms. avg: 1.54 ± 0.00ms min: 0.09 ± 0.00ms max: 5.25 ± 0.00ms p99: 3.78 ± 0.00ms Peak RAM: 83.500MB
random_route_mld	1000 routes, 3 coordinates, no alternatives, overview=full, steps=true ops: 146.74 ± 0.02 ops/s. best: 146.76ops/s. total: 6705.87 ± 1.05ms. best: 6704.68ms. avg: 6.81 ± 0.00ms min: 0.42 ± 0.00ms max: 16.37 ± 0.02ms p99: 11.08 ± 0.01ms 1000 routes, 2 coordinates, 3 alternatives, overview=full, steps=true ops: 139.17 ± 0.02 ops/s. best: 139.20ops/s. total: 7185.68 ± 0.78ms. best: 7184.05ms. avg: 7.19 ± 0.00ms min: 0.11 ± 0.00ms max: 16.54 ± 0.08ms p99: 15.57 ± 0.02ms 1000 routes, 3 coordinates, no alternatives, overview=false, steps=false ops: 204.62 ± 0.03 ops/s. best: 204.67ops/s. total: 4809.01 ± 0.81ms. best: 4807.68ms. avg: 4.89 ± 0.00ms min: 0.36 ± 0.00ms max: 13.58 ± 0.03ms p99: 8.43 ± 0.01ms 1000 routes, 2 coordinates, 3 alternatives, overview=false, steps=false ops: 175.00 ± 0.04 ops/s. best: 175.06ops/s. total: 5714.13 ± 1.20ms. best: 5712.32ms. avg: 5.71 ± 0.00ms min: 0.08 ± 0.00ms max: 12.83 ± 0.02ms p99: 11.99 ± 0.02ms Peak RAM: 73.531MB	1000 routes, 3 coordinates, no alternatives, overview=full, steps=true ops: 168.21 ± 0.04 ops/s. best: 168.28ops/s. total: 5850.00 ± 1.26ms. best: 5847.43ms. avg: 5.95 ± 0.00ms min: 0.42 ± 0.00ms max: 13.64 ± 0.03ms p99: 9.53 ± 0.02ms 1000 routes, 2 coordinates, 3 alternatives, overview=full, steps=true ops: 161.19 ± 0.01 ops/s. best: 161.20ops/s. total: 6204.02 ± 0.48ms. best: 6203.31ms. avg: 6.20 ± 0.00ms min: 0.11 ± 0.00ms max: 14.53 ± 0.06ms p99: 13.57 ± 0.01ms 1000 routes, 3 coordinates, no alternatives, overview=false, steps=false ops: 248.02 ± 0.03 ops/s. best: 248.06ops/s. total: 3967.49 ± 0.55ms. best: 3966.74ms. avg: 4.03 ± 0.00ms min: 0.35 ± 0.00ms max: 10.85 ± 0.01ms p99: 6.92 ± 0.01ms 1000 routes, 2 coordinates, 3 alternatives, overview=false, steps=false ops: 210.60 ± 0.03 ops/s. best: 210.65ops/s. total: 4748.30 ± 0.64ms. best: 4747.30ms. avg: 4.75 ± 0.00ms min: 0.08 ± 0.00ms max: 11.03 ± 0.02ms p99: 10.06 ± 0.02ms Peak RAM: 73.531MB
random_table_ch	250 tables, 3 coordinates ops: 1158.95 ± 5.04 ops/s. best: 1162.61ops/s. total: 215.72 ± 0.94ms. best: 215.03ms. avg: 0.86 ± 0.00ms min: 0.60 ± 0.00ms max: 1.26 ± 0.14ms p99: 1.11 ± 0.02ms 250 tables, 25 coordinates ops: 128.52 ± 0.02 ops/s. best: 128.56ops/s. total: 1945.21 ± 0.29ms. best: 1944.57ms. avg: 7.78 ± 0.00ms min: 7.09 ± 0.01ms max: 8.53 ± 0.01ms p99: 8.44 ± 0.01ms 250 tables, 50 coordinates ops: 62.40 ± 0.01 ops/s. best: 62.41ops/s. total: 4006.70 ± 0.74ms. best: 4005.48ms. avg: 16.03 ± 0.00ms min: 14.88 ± 0.02ms max: 17.14 ± 0.01ms p99: 16.99 ± 0.02ms Peak RAM: 62.500MB	250 tables, 3 coordinates ops: 1248.62 ± 5.26 ops/s. best: 1253.29ops/s. total: 200.23 ± 0.85ms. best: 199.48ms. avg: 0.80 ± 0.00ms min: 0.55 ± 0.00ms max: 1.16 ± 0.12ms p99: 1.04 ± 0.02ms 250 tables, 25 coordinates ops: 137.83 ± 0.01 ops/s. best: 137.85ops/s. total: 1813.80 ± 0.18ms. best: 1813.51ms. avg: 7.26 ± 0.00ms min: 6.58 ± 0.01ms max: 7.99 ± 0.01ms p99: 7.86 ± 0.01ms 250 tables, 50 coordinates ops: 66.83 ± 0.00 ops/s. best: 66.84ops/s. total: 3740.57 ± 0.22ms. best: 3740.23ms. avg: 14.96 ± 0.00ms min: 13.86 ± 0.01ms max: 15.99 ± 0.01ms p99: 15.84 ± 0.01ms Peak RAM: 63.000MB
random_table_mld	250 tables, 3 coordinates ops: 227.05 ± 0.28 ops/s. best: 227.45ops/s. total: 1101.09 ± 1.35ms. best: 1099.14ms. avg: 4.40 ± 0.01ms min: 3.46 ± 0.00ms max: 5.67 ± 0.01ms p99: 5.47 ± 0.05ms 250 tables, 25 coordinates ops: 23.83 ± 0.00 ops/s. best: 23.84ops/s. total: 10490.59 ± 2.20ms. best: 10487.77ms. avg: 41.96 ± 0.01ms min: 38.70 ± 0.04ms max: 45.83 ± 0.11ms p99: 45.19 ± 0.05ms 250 tables, 50 coordinates ops: 11.02 ± 0.00 ops/s. best: 11.02ops/s. total: 22690.88 ± 7.06ms. best: 22675.92ms. avg: 90.76 ± 0.03ms min: 86.23 ± 0.06ms max: 96.50 ± 0.10ms p99: 94.81 ± 0.07ms Peak RAM: 62.734MB	250 tables, 3 coordinates ops: 264.79 ± 0.35 ops/s. best: 265.04ops/s. total: 944.16 ± 1.25ms. best: 943.26ms. avg: 3.78 ± 0.00ms min: 2.94 ± 0.00ms max: 4.86 ± 0.02ms p99: 4.70 ± 0.04ms 250 tables, 25 coordinates ops: 27.37 ± 0.00 ops/s. best: 27.38ops/s. total: 9132.52 ± 0.65ms. best: 9131.27ms. avg: 36.53 ± 0.00ms min: 33.65 ± 0.02ms max: 39.92 ± 0.02ms p99: 39.34 ± 0.03ms 250 tables, 50 coordinates ops: 12.58 ± 0.00 ops/s. best: 12.58ops/s. total: 19880.20 ± 2.43ms. best: 19876.75ms. avg: 79.52 ± 0.01ms min: 75.39 ± 0.04ms max: 84.58 ± 0.02ms p99: 83.10 ± 0.08ms Peak RAM: 62.703MB
random_trip_ch	250 trips, 3 coordinates ops: 329.19 ± 0.48 ops/s. best: 329.57ops/s. total: 759.44 ± 1.11ms. best: 758.56ms. avg: 3.04 ± 0.00ms min: 1.59 ± 0.01ms max: 4.46 ± 0.00ms p99: 4.23 ± 0.01ms 250 trips, 5 coordinates ops: 214.38 ± 0.05 ops/s. best: 214.46ops/s. total: 1166.18 ± 0.26ms. best: 1165.72ms. avg: 4.66 ± 0.00ms min: 2.93 ± 0.01ms max: 5.99 ± 0.01ms p99: 5.86 ± 0.01ms Peak RAM: 73.000MB	250 trips, 3 coordinates ops: 345.60 ± 0.46 ops/s. best: 345.93ops/s. total: 723.37 ± 0.96ms. best: 722.69ms. avg: 2.89 ± 0.00ms min: 1.50 ± 0.01ms max: 4.28 ± 0.01ms p99: 4.05 ± 0.01ms 250 trips, 5 coordinates ops: 223.73 ± 0.07 ops/s. best: 223.79ops/s. total: 1117.44 ± 0.35ms. best: 1117.11ms. avg: 4.47 ± 0.00ms min: 2.76 ± 0.00ms max: 5.81 ± 0.01ms p99: 5.62 ± 0.01ms Peak RAM: 73.000MB
random_trip_mld	250 trips, 3 coordinates ops: 106.72 ± 0.05 ops/s. best: 106.77ops/s. total: 2342.57 ± 1.14ms. best: 2341.47ms. avg: 9.37 ± 0.00ms min: 5.53 ± 0.01ms max: 12.12 ± 0.01ms p99: 12.05 ± 0.01ms 250 trips, 5 coordinates ops: 68.72 ± 0.01 ops/s. best: 68.74ops/s. total: 3637.85 ± 0.46ms. best: 3637.14ms. avg: 14.55 ± 0.00ms min: 10.08 ± 0.01ms max: 17.91 ± 0.03ms p99: 17.46 ± 0.02ms Peak RAM: 68.500MB	250 trips, 3 coordinates ops: 124.13 ± 0.07 ops/s. best: 124.18ops/s. total: 2014.07 ± 1.09ms. best: 2013.25ms. avg: 8.06 ± 0.00ms min: 4.77 ± 0.01ms max: 10.46 ± 0.02ms p99: 10.41 ± 0.01ms 250 trips, 5 coordinates ops: 79.66 ± 0.02 ops/s. best: 79.68ops/s. total: 3138.40 ± 0.65ms. best: 3137.42ms. avg: 12.55 ± 0.00ms min: 8.65 ± 0.01ms max: 15.50 ± 0.04ms p99: 15.10 ± 0.02ms Peak RAM: 68.500MB
route_ch	1000 routes, 3 coordinates, no alternatives, overview=full, steps=true 635.942ms 0.635942ms/req 1000 routes, 2 coordinates, 3 alternatives, overview=full, steps=true 778.264ms 0.778264ms/req 1000 routes, 3 coordinates, no alternatives, overview=false, steps=false 245.431ms 0.245431ms/req 1000 routes, 2 coordinates, 3 alternatives, overview=false, steps=false 214.527ms 0.214527ms/req	1000 routes, 3 coordinates, no alternatives, overview=full, steps=true 627.748ms 0.627748ms/req 1000 routes, 2 coordinates, 3 alternatives, overview=full, steps=true 773.678ms 0.773678ms/req 1000 routes, 3 coordinates, no alternatives, overview=false, steps=false 238.52ms 0.23852ms/req 1000 routes, 2 coordinates, 3 alternatives, overview=false, steps=false 211.042ms 0.211042ms/req
route_mld	1000 routes, 3 coordinates, no alternatives, overview=full, steps=true 806.896ms 0.806896ms/req 1000 routes, 2 coordinates, 3 alternatives, overview=full, steps=true 1038.03ms 1.03803ms/req 1000 routes, 3 coordinates, no alternatives, overview=false, steps=false 413.451ms 0.413451ms/req 1000 routes, 2 coordinates, 3 alternatives, overview=false, steps=false 460.871ms 0.460871ms/req	1000 routes, 3 coordinates, no alternatives, overview=full, steps=true 749.022ms 0.749022ms/req 1000 routes, 2 coordinates, 3 alternatives, overview=full, steps=true 974.73ms 0.97473ms/req 1000 routes, 3 coordinates, no alternatives, overview=false, steps=false 355.375ms 0.355375ms/req 1000 routes, 2 coordinates, 3 alternatives, overview=false, steps=false 397.201ms 0.397201ms/req
rtree	1 result: 226.994ms -> 0.0226994 ms/query 10 results: 257.736ms -> 0.0257736 ms/query	1 result: 227.271ms -> 0.0227271 ms/query 10 results: 258.023ms -> 0.0258023 ms/query

DennisOSRM

Great find! I'd love to see a simpler implementation that keeps heap logic simple and contained within the implementation. Left some specific comments at the code site.

unit_tests/CMakeLists.txt

DennisOSRM · 2024-08-22T20:01:56Z

include/util/d_ary_heap.hpp

+
+namespace osrm::util
+{
+template <typename HeapData, int Arity, typename Comparator = std::less<HeapData>> class DAryHeap


Having the reorder handler as a template parameter seems somewhat confusing.

Do have a need to specify multiple ones at all? If this is needed, should the be implemented inside the heap?

Can we possibly aim for a simpler implementation of this?

Having the reorder handler as a template parameter seems somewhat confusing.

Well, the intention here was to be able to pass lambda and avoid using std::function which has certain performance impact... Such template parameters is frequent pattern for such things - if you want to pass lambda, compiler will infer type of this lambda for you in a most efficient way. At the same time I agree that probably having the fact that we pass only single type of function in there worth trying to make this more obvious (not sure it will be simple though, but I'll try 😄 )

If this is needed, should the be implemented inside the heap?

What do you mean by "inside the heap"? In my initial implementation DAryHeap wasn't a separate class and all code from here "lived" directly in QueryHeap, but I found it a bit messy and decided to decouple things a bit to keep code structure a bit similar to what we have now (QueryHeap and separate boost::d_ary_heap before and QueryHeap and separate DAryHeap after). I could try to return back to initial implementation...

WDYT?

Or you mean that implementation of reorderHandler should be a part of DAryHeap? I am not sure it is also doable without merging DAryHeap to QueryHeap - this reorderHandler is kind of a "glue" between DAryHeap and QueryHeap...

Another option I see is... make it super straightforward. DAryHeap now is kind of generic DAryHeap implementation, but we could "couple" it to our needs, i.e. just take reference to inserted_nodes from QueryHeap and do all things we are currently doing in reorderHandler directly on that vector. Should also work without issues as well.

Anyway @DennisOSRM would love to hear your opinion on this please 😄

SO question about passing lambdas as parameters https://stackoverflow.com/questions/6458612/proper-way-to-receive-a-lambda-as-parameter-by-reference

And one more blogpost with concrete example of why template can be more preferable here… https://wolchok.org/posts/cxx-trap-2-std-function/

@DennisOSRM WDYT? Tbh I am not sure I have good ideas how to remove template parameter here - all ideas have their own drawbacks. May be you have some? Or may be worth introducing a bit of complexity in favor of performance? 😀

SiarheiFedartsou changed the title ~~Prototype using custom d-ary heap implementation~~ Use custom d-ary heap implementation Aug 20, 2024

SiarheiFedartsou marked this pull request as ready for review August 22, 2024 15:08

SiarheiFedartsou requested review from DennisOSRM and mjjbell August 22, 2024 15:08

DennisOSRM reviewed Aug 22, 2024

View reviewed changes

Use custom d-ary heap implementation

eb50311

SiarheiFedartsou force-pushed the sf-query-heap branch from 4b3dbc3 to eb50311 Compare August 22, 2024 20:16

SiarheiFedartsou requested a review from DennisOSRM August 26, 2024 19:49

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use custom d-ary heap implementation #7017

Use custom d-ary heap implementation #7017

SiarheiFedartsou commented Aug 19, 2024 •

edited by github-actions bot

Loading

DennisOSRM left a comment

DennisOSRM Aug 22, 2024

SiarheiFedartsou Aug 22, 2024

SiarheiFedartsou Aug 22, 2024

SiarheiFedartsou Aug 22, 2024

SiarheiFedartsou Aug 22, 2024

SiarheiFedartsou Aug 24, 2024

SiarheiFedartsou Aug 24, 2024

Use custom d-ary heap implementation #7017

Are you sure you want to change the base?

Use custom d-ary heap implementation #7017

Conversation

SiarheiFedartsou commented Aug 19, 2024 • edited by github-actions bot Loading

Context

Benchmark highlights

Description

Benchmark Results

DennisOSRM left a comment

Choose a reason for hiding this comment

DennisOSRM Aug 22, 2024

Choose a reason for hiding this comment

SiarheiFedartsou Aug 22, 2024

Choose a reason for hiding this comment

SiarheiFedartsou Aug 22, 2024

Choose a reason for hiding this comment

SiarheiFedartsou Aug 22, 2024

Choose a reason for hiding this comment

SiarheiFedartsou Aug 22, 2024

Choose a reason for hiding this comment

SiarheiFedartsou Aug 24, 2024

Choose a reason for hiding this comment

SiarheiFedartsou Aug 24, 2024

Choose a reason for hiding this comment

SiarheiFedartsou commented Aug 19, 2024 •

edited by github-actions bot

Loading