TEST(SumNode, GenerateAndSumHaveEqualAmount2) {
std::function<void(Context&)> start_func =
[](Context& ctx) {
// TODO(ms): Replace this with some test-specific rendered file
auto input = ReadLines(ctx, "inputs/test1")
.Map([](const std::string& line) {
return std::stoi(line);
});
auto ones = input.Map([](int in) {
return in;
});
auto add_function = [](int in1, int in2) {
return in1 + in2;
};
DIA<int> coll = ones.Collapse();
ASSERT_EQ(136, coll.Sum(add_function));
ASSERT_EQ(16u, coll.Size());
};
api::RunLocalTests(start_func);
}
TEST(Stage, CountReferencesLOpNode) {
std::function<void(Context&)> start_func =
[](Context& ctx) {
auto integers = Generate(
ctx,
[](const size_t& index) {
return static_cast<int>(index) + 1;
},
16);
auto duplicate_elements = [](int in, auto emit) {
emit(in);
emit(in);
};
auto modulo_two = [](int in) {
return (in % 2);
};
auto add_function = [](int in1, int in2) {
return in1 + in2;
};
// Create a new DIA references to Generate
auto doubles = integers.FlatMap(duplicate_elements);
// Create a child references to Generate
// Create a new DIA reference to LOpNode
DIA<int> quadruples = integers.FlatMap(duplicate_elements).Cache();
// Create new child reference to LOpNode
auto reduced = quadruples.ReduceBy(modulo_two, add_function);
// Trigger execution
std::vector<int> out_vec = reduced.AllGather();
// 2x DIA reference + 1x child reference
ASSERT_EQ(integers.node_refcount(), 3u);
ASSERT_EQ(doubles.node_refcount(), 3u);
// 1x DIA reference + 1x child reference
ASSERT_EQ(quadruples.node_refcount(), 2u);
// 1x DIA reference + 0x child reference
ASSERT_EQ(reduced.node_refcount(), 1u);
};
api::RunLocalTests(start_func);
}
std::pair<ValueType, ValueType>
PickPivots(const DIA<ValueType, InStack>& data, size_t size, size_t rank,
const Compare& compare = Compare()) {
api::Context& ctx = data.context();
const size_t num_workers(ctx.num_workers());
const double size_d = static_cast<double>(size);
const double p = 20 * sqrt(static_cast<double>(num_workers)) / size_d;
// materialized at worker 0
auto sample = data.BernoulliSample(p).Gather();
std::pair<ValueType, ValueType> pivots;
if (ctx.my_rank() == 0) {
LOG << "got " << sample.size() << " samples (p = " << p << ")";
// Sort the samples
std::sort(sample.begin(), sample.end(), compare);
const double base_pos =
static_cast<double>(rank * sample.size()) / size_d;
const double offset = pow(size_d, 0.25 + delta);
long lower_pos = static_cast<long>(floor(base_pos - offset));
long upper_pos = static_cast<long>(floor(base_pos + offset));
size_t lower = static_cast<size_t>(std::max(0L, lower_pos));
size_t upper = static_cast<size_t>(
std::min(upper_pos, static_cast<long>(sample.size() - 1)));
assert(0 <= lower && lower < sample.size());
assert(0 <= upper && upper < sample.size());
LOG << "Selected pivots at positions " << lower << " and " << upper
<< ": " << sample[lower] << " and " << sample[upper];
pivots = std::make_pair(sample[lower], sample[upper]);
}
pivots = ctx.net.Broadcast(pivots);
LOGM << "pivots: " << pivots.first << " and " << pivots.second;
return pivots;
}
static DIA<ValueType> MakeCollapse(const DIA<ValueType, Stack>& dia) {
assert(dia.IsValid());
// Create new CollapseNode. Transfer stack from rhs to
// CollapseNode. Build new DIA with empty stack and CollapseNode
using CollapseNode = api::CollapseNode<ValueType>;
return DIA<ValueType>(tlx::make_counting<CollapseNode>(dia));
}
TEST(Graph, WhileLoop) {
std::function<void(Context&)> start_func =
[](Context& ctx) {
auto integers = Generate(
ctx,
[](const size_t& index) -> size_t {
return index;
},
16);
auto flatmap_duplicate = [](size_t in, auto emit) {
emit(in);
emit(in);
};
auto map_multiply = [](size_t in) {
return 2 * in;
};
DIA<size_t> squares = integers.Collapse();
size_t sum = 0;
// run loop four times, inflating DIA of 16 items -> 256
while (sum < 64) {
auto pairs = squares.FlatMap(flatmap_duplicate);
auto multiplied = pairs.Map(map_multiply);
squares = multiplied.Cache();
sum = squares.Size();
}
std::vector<size_t> out_vec = squares.AllGather();
ASSERT_EQ(64u, out_vec.size());
ASSERT_EQ(64u, squares.Size());
ctx.stats_graph().BuildLayout("loop.out");
};
api::RunLocalTests(start_func);
}
DIA<ValueType, Stack>::DIA(const DIA<ValueType, AnyStack>& rhs) {
// Create new CollapseNode. Transfer stack from rhs to CollapseNode. Build
// new DIA with empty stack and CollapseNode
using CollapseNode =
api::CollapseNode<ValueType, DIA<ValueType, AnyStack> >;
LOG0 << "WARNING: cast to DIA creates CollapseNode instead of inline chaining.";
LOG0 << "Consider whether you can use auto instead of DIA.";
StatsNode* stats_node = rhs.AddChildStatsNode("Collapse", DIANodeType::COLLAPSE);
node_ = std::make_shared<CollapseNode>(rhs, stats_node);
// stack_ is default constructed.
stats_parents_.emplace_back(stats_node);
}
void OutputSVG(const std::string& svg_path, double svg_scale,
const DIA<Point<2> >& point_dia,
const KMeansModel<Point<2> >& model) {
double width = 0, height = 0;
using Point2D = Point<2>;
const std::vector<Point2D>& centroids = model.centroids();
std::vector<PointClusterId<Point2D> > list =
model.ClassifyPairs(point_dia).Gather();
for (const PointClusterId<Point2D>& p : list) {
width = std::max(width, p.first.x[0]);
height = std::max(height, p.first.x[1]);
}
if (point_dia.context().my_rank() != 0) return;
std::ofstream os(svg_path);
os << "<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"no\"?>\n";
os << "<svg\n";
os << " xmlns:dc=\"http://purl.org/dc/elements/1.1/\"\n";
os << " xmlns:cc=\"http://creativecommons.org/ns#\"\n";
os << " xmlns:rdf=\"http://www.w3.org/1999/02/22-rdf-syntax-ns#\"\n";
os << " xmlns:svg=\"http://www.w3.org/2000/svg\"\n";
os << " xmlns=\"http://www.w3.org/2000/svg\"\n";
os << " version=\"1.1\" id=\"svg2\" width=\"" << width * svg_scale
<< "\" height=\"" << height * svg_scale << "\">\n";
os << " <g id=\"layer1\">\n";
for (const PointClusterId<Point2D>& p : list) {
os << " <circle r=\"1\" cx=\"" << p.first.x[0] * svg_scale
<< "\" cy=\"" << p.first.x[1] * svg_scale
<< "\" style=\"stroke:none;stroke-opacity:1;fill:"
<< SVGColor(p.second) << ";fill-opacity:1\" />\n";
}
for (size_t i = 0; i < centroids.size(); ++i) {
const Point2D& p = centroids[i];
os << " <circle r=\"4\" cx=\"" << p.x[0] * svg_scale
<< "\" cy=\"" << p.x[1] * svg_scale
<< "\" style=\"stroke:black;stroke-opacity:1;fill:"
<< SVGColor(i) << ";fill-opacity:1\" />\n";
}
os << " </g>\n";
os << "</svg>\n";
}
TEST(ZipNode, TwoDisbalancedStringArrays) {
// first DIA is heavily balanced to the first workers, second DIA is
// balanced to the last workers.
std::function<void(Context&)> start_func =
[](Context& ctx) {
// generate random strings with 10..20 characters
auto input_gen = Generate(
ctx,
[](size_t index) -> std::string {
std::default_random_engine rng(
123456 + static_cast<unsigned>(index));
std::uniform_int_distribution<size_t> length(10, 20);
rng(); // skip one number
return common::RandomString(
length(rng), rng, "abcdefghijklmnopqrstuvwxyz")
+ std::to_string(index);
},
test_size);
DIA<std::string> input = input_gen.Cache();
std::vector<std::string> vinput = input.AllGather();
ASSERT_EQ(test_size, vinput.size());
// Filter out strings that start with a-e
auto input1 = input.Filter(
[](const std::string& str) { return str[0] <= 'e'; });
// Filter out strings that start with w-z
auto input2 = input.Filter(
[](const std::string& str) { return str[0] >= 'w'; });
// zip
auto zip_result = input1.Zip(
input2, [](const std::string& a, const std::string& b) {
return a + b;
});
// check result
std::vector<std::string> res = zip_result.AllGather();
// recalculate result locally
std::vector<std::string> check;
{
std::vector<std::string> v1, v2;
for (size_t index = 0; index < vinput.size(); ++index) {
const std::string& s1 = vinput[index];
if (s1[0] <= 'e') v1.push_back(s1);
if (s1[0] >= 'w') v2.push_back(s1);
}
ASSERT_EQ(v1, input1.AllGather());
ASSERT_EQ(v2, input2.AllGather());
for (size_t i = 0; i < std::min(v1.size(), v2.size()); ++i) {
check.push_back(v1[i] + v2[i]);
// sLOG1 << check.back();
}
}
for (size_t i = 0; i != res.size(); ++i) {
sLOG0 << res[i] << " " << check[i] << (res[i] == check[i]);
}
ASSERT_EQ(check.size(), res.size());
ASSERT_EQ(check, res);
};
api::RunLocalTests(start_func);
}
auto PageRank(const DIA<std::string, InStack>&in, api::Context & ctx, int iter) {
DIA<Page_Link> input = in.Map(
[](const std::string& input) {
auto split = thrill::common::Split(input, " ");
LOG0 << "input "
<< (std::stoi(split[0]) - 1)
<< " "
<< (std::stoi(split[1]) - 1);
// set base of page_id to 0
return Page_Link((size_t)(std::stoi(split[0]) - 1),
(size_t)(std::stoi(split[1]) - 1));
}).Cache(); // TODO(SL): when Cache() is removed, code doesn't compile,
// auto cannot be used either
// aggregate all outgoing links of a page in this format:
//
// URL OUTGOING
// ([linked_url, linked_url, ...])
// ([linked_url, linked_url, ...])
// ([linked_url, linked_url, ...])
// ...
// get number of nodes by finding max page_id
// add 1 to max node_id to get number of nodes because of node_id 0
const auto number_nodes = input.Sum(
[](const Page_Link& in1, const Page_Link& in2) {
Node first = std::max(in1.first, in2.first);
Node second = std::max(in1.second, in2.second);
return std::make_pair(std::max(first, second), first);
}).first + 1;
LOG << "number_nodes " << number_nodes;
// group outgoing links
auto links = input.GroupByIndex<Outgoings>(
[](Page_Link p) { return p.first; },
[](auto& r, Key) {
std::vector<Node> all;
while (r.HasNext()) {
all.push_back(r.Next().second);
}
// std::string s = "{";
// for (auto e : all) {
// s+= std::to_string(e) + ", ";
// }
// LOG << "links " << s << "}";
return all;
},
number_nodes).Cache();
// initialize all ranks to 1.0
//
// (url, rank)
// (url, rank)
// (url, rank)
// ...
// auto ranks = Generate(ctx, [](const size_t& index) {
// return std::make_pair(index, 1.0);
// }, number_nodes).Cache();
auto ranks = Generate(ctx,
[](const size_t&) {
return (Rank)1.0;
}, number_nodes).Cache();
auto node_ids = Generate(ctx,
[](const size_t& index) {
return index + 1;
}, number_nodes);
// do iterations
for (int i = 0; i < iter; ++i) {
LOG << "iteration " << i;
// for all outgoing link, get their rank contribution from all
// links by doing:
//
// 1) group all outgoing links with rank of its parent page: (Zip)
//
// ([linked_url, linked_url, ...], rank_parent)
// ([linked_url, linked_url, ...], rank_parent)
// ([linked_url, linked_url, ...], rank_parent)
//
// 2) compute rank contribution for each linked_url: (FlatMap)
//
// (linked_url, rank / OUTGOING.size)
// (linked_url, rank / OUTGOING.size)
// (linked_url, rank / OUTGOING.size)
// ...
std::cout << links.Size() << std::endl;
std::cout << ranks.Size() << std::endl;
assert(links.Size() == ranks.Size());
// TODO(SL): when Zip/FlatMap chained, code doesn't compile, please check
DIA<Outgoings_Rank> outs_rank = links.Zip(ranks,
[](const Outgoings& l, const Rank r) {
// std::string s = "{";
// for (auto e : l) {
// s += std::to_string(e) + ", ";
// }
// s += "}";
// LOG << "contribs1 " << s << " " << r;
return std::make_pair(l, r);
});
auto contribs = outs_rank.FlatMap<Page_Rank>(
[](const Outgoings_Rank& p, auto emit) {
if (p.first.size() > 0) {
Rank rank_contrib = p.second / p.first.size();
// assert (rank_contrib <= 1);
for (auto e : p.first) {
LOG << "contribs2 " << e << " " << rank_contrib;
emit(std::make_pair(e, rank_contrib));
}
}
});
// reduce all rank contributions by adding all rank contributions
// and compute the new rank with 0.15 * 0.85 * sum_rank_contribs
//
// (url, rank)
// (url, rank)
// (url, rank)
// ...
// auto sum_rank_contrib_fn = [](const Page_Rank& p1, const Page_Rank& p2) {
// assert(p1.first == p2.first);
// return p1.second + p2.second;
// };
ranks = contribs.ReduceToIndex(
[](const Page_Rank& p) { return p.first; },
[](const Page_Rank& p1, const Page_Rank& p2) {
return std::make_pair(p1.first, p1.second + p2.second);
}, number_nodes)
.Map(
[](const Page_Rank p) {
LOG << "ranks2 in " << p.first << "-" << p.second;
if (std::fabs(p.second) <= 1E-5) {
LOG << "ranks2 " << 0.0;
return (Rank)0.0;
}
else {
LOG << "ranks2 " << f + s * p.second;
return f + s * p.second;
}
}).Keep().Collapse();
}
// write result to line. add 1 to node_ids to revert back to normal
auto res = ranks.Zip(node_ids,
[](const Rank r, const Node n) {
return std::to_string(n)
+ ": " + std::to_string(r);
});
assert(res.Size() == links.Size());
return res;
}
void OutputSVG(const std::string& svg_path, double svg_scale,
const DIA<DataPoint<Vector> >& point_dia,
const Vector& model) {
double width = 0, height = 0, min_vert = 0, max_vert = 0, min_hor = 0, max_hor = 0;
std::vector<DataPoint<Vector> > list = point_dia.Gather();
for (const DataPoint<Vector>& p : list) {
min_hor = std::min(min_hor, p.data.x[0]);
max_hor = std::max(max_hor, p.data.x[0]);
min_vert = std::min(min_vert, p.label);
max_vert = std::max(max_vert, p.label);
}
double weight = model.x[0];
double y1 = min_hor * weight;
double y2 = max_hor * weight;
min_vert = std::min(min_vert, y1);
min_vert = std::min(min_vert, y2);
max_vert = std::max(max_vert, y1);
max_vert = std::max(max_vert, y2);
width = max_hor - min_hor;
height = max_vert - min_vert;
std::ofstream os(svg_path);
os << "<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"no\"?>\n";
os << "<svg\n";
os << " xmlns:dc=\"http://purl.org/dc/elements/1.1/\"\n";
os << " xmlns:cc=\"http://creativecommons.org/ns#\"\n";
os << " xmlns:rdf=\"http://www.w3.org/1999/02/22-rdf-syntax-ns#\"\n";
os << " xmlns:svg=\"http://www.w3.org/2000/svg\"\n";
os << " xmlns=\"http://www.w3.org/2000/svg\"\n";
os << " version=\"1.1\" id=\"svg2\" width=\"" << width * svg_scale
<< "\" height=\"" << height * svg_scale << "\">\n";
os << " <g id=\"layer1\">\n";
// Draw grid
os << " <line x1=\"0\" y1=\"" << (height + min_vert) * svg_scale
<< "\" x2=\"" << width * svg_scale << "\" y2=\"" << (height + min_vert) * svg_scale
<< "\" stroke-width=\"1\" stroke=\"#777777\" style=\"stroke-opacity:0.3\" />\n";
os << " <line x1=\"" << -min_hor * svg_scale << "\" y1=\"0\""
<< " x2=\"" << -min_hor * svg_scale << "\" y2=\"" << height * svg_scale
<< "\" stroke-width=\"1\" stroke=\"#777777\" style=\"stroke-opacity:0.3\" />\n";
// Draw points
for (const DataPoint<Vector>& p : list) {
os << " <circle r=\"1\" cx=\"" << (p.data.x[0] - min_hor) * svg_scale
<< "\" cy=\"" << (height - p.label + min_vert) * svg_scale
<< "\" style=\"stroke:none;stroke-opacity:1;fill:#45a2d1;fill-opacity:1\" />\n";
}
// Draw line
os << " <line x1=\"0\" y1=\"" << (height - y1 + min_vert) * svg_scale
<< "\" x2=\"" << width * svg_scale << "\" y2=\"" << (height - y2 + min_vert) * svg_scale
<< "\" stroke-width=\"1\" stroke=\"#ff9900\" />\n";
os << " </g>\n";
os << "</svg>\n";
}
ValueType Select(const DIA<ValueType, InStack>& data, size_t rank,
const Compare& compare = Compare()) {
api::Context& ctx = data.context();
const size_t size = data.Size();
assert(0 <= rank && rank < size);
if (size <= base_case_size) {
// base case, gather all data at worker with rank 0
ValueType result = ValueType();
auto elements = data.Gather();
if (ctx.my_rank() == 0) {
assert(rank < elements.size());
std::nth_element(elements.begin(), elements.begin() + rank,
elements.end(), compare);
result = elements[rank];
LOG << "base case: " << size << " elements remaining, result is "
<< result;
}
result = ctx.net.Broadcast(result);
return result;
}
ValueType left_pivot, right_pivot;
std::tie(left_pivot, right_pivot) = PickPivots(data, size, rank, compare);
size_t left_size, middle_size, right_size;
using PartSizes = std::pair<size_t, size_t>;
std::tie(left_size, middle_size) =
data.Map(
[&](const ValueType& elem) -> PartSizes {
if (compare(elem, left_pivot))
return PartSizes { 1, 0 };
else if (!compare(right_pivot, elem))
return PartSizes { 0, 1 };
else
return PartSizes { 0, 0 };
})
.Sum(
[](const PartSizes& a, const PartSizes& b) -> PartSizes {
return PartSizes { a.first + b.first, a.second + b.second };
},
PartSizes { 0, 0 });
right_size = size - left_size - middle_size;
LOGM << "left_size = " << left_size << ", middle_size = " << middle_size
<< ", right_size = " << right_size << ", rank = " << rank;
if (rank == left_size) {
// all the elements strictly smaller than the left pivot are on the left
// side -> left_size-th element is the left pivot
LOGM << "result is left pivot: " << left_pivot;
return left_pivot;
}
else if (rank == left_size + middle_size - 1) {
// only the elements strictly greater than the right pivot are on the
// right side, so the result is the right pivot in this case
LOGM << "result is right pivot: " << right_pivot;
return right_pivot;
}
else if (rank < left_size) {
// recurse on the left partition
LOGM << "Recursing left, " << left_size
<< " elements remaining (rank = " << rank << ")\n";
auto left = data.Filter(
[&](const ValueType& elem) -> bool {
return compare(elem, left_pivot);
}).Collapse();
assert(left.Size() == left_size);
return Select(left, rank, compare);
}
else if (left_size + middle_size <= rank) {
// recurse on the right partition
LOGM << "Recursing right, " << right_size
<< " elements remaining (rank = " << rank - left_size - middle_size
<< ")\n";
auto right = data.Filter(
[&](const ValueType& elem) -> bool {
return compare(right_pivot, elem);
}).Collapse();
assert(right.Size() == right_size);
return Select(right, rank - left_size - middle_size, compare);
}
else {
// recurse on the middle partition
LOGM << "Recursing middle, " << middle_size
<< " elements remaining (rank = " << rank - left_size << ")\n";
auto middle = data.Filter(
[&](const ValueType& elem) -> bool {
return !compare(elem, left_pivot) &&
!compare(right_pivot, elem);
}).Collapse();
assert(middle.Size() == middle_size);
return Select(middle, rank - left_size, compare);
}
}
