school
/
csci5451
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

assignment 3

This commit is contained in:
Michael Zhang 2023-12-10 16:58:28 -06:00
parent d75da8de6d
commit 442638205c
9 changed files with 261 additions and 219 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
Dockerfile
View File

@ -22,6 +22,7 @@ RUN apt update -y && apt install -y --no-install-recommends \
pkg-config \
python3 \
python3-pip \
python3-venv \
texlive-latex-base \
texlive-latex-extra \
valgrind \

3
assignments/03/.envrc Normal file
View File

@ -0,0 +1,3 @@
layout python3
export OMPI_ALLOW_RUN_AS_ROOT=1
export OMPI_ALLOW_RUN_AS_ROOT_CONFIRM=1

5
assignments/03/.gitignore vendored
View File

@ -3,4 +3,7 @@ compile_commands.json
.cache
report.pdf
*.tar.gz
out.txt
out.txt
dataset/gen_*.txt
.direnv

2
assignments/03/Makefile
View File

@ -16,7 +16,7 @@ run:
report.pdf: report.typ
typst compile $< $@
zhan4854.tar.gz: Makefile ASSIGNMENT.md lpa.cpp report.pdf
zhan4854.tar.gz: Makefile ASSIGNMENT.md lpa.cpp report.pdf dataset/gen2.py
mkdir -p zhan4854
cp $^ zhan4854
tar -czvf $@ zhan4854

3
assignments/03/bench.sh
View File

@ -2,7 +2,6 @@ for dataset in $(echo "1000.txt" "10000.txt" "100000.txt" "1000000.txt"); do
for processors in $(echo 1 2 4 8 16 | tr ' ' '\n'); do
# file="dataset/both_$dataset"
file="/export/scratch/CSCI5451_F23/assignment-3/dataset/$dataset"
echo $processors $file;
mpirun -n $processors ./lpa $file graphout.txt >> stdout.txt
mpirun -n $processors ./lpa $file "graph_out/$dataset-$processors.txt" >> "stdout_out/$dataset-$processors.txt"
done
done

25
assignments/03/dataset/gen2.py Normal file
View File

@ -0,0 +1,25 @@
import igraph as ig
import random
import sys
try:
N = int(sys.argv[1])
except:
N = 1000
random.seed(0)
g = ig.Graph.Growing_Random(N, 5)
components = g.connected_components(mode='weak')
print(len(components))
with open(f"dataset/gen_{N}.txt", "w") as f:
both_edges = []
for edge in g.es:
both_edges.append((edge.source, edge.target))
both_edges.append((edge.target, edge.source))
num_edges = len(both_edges)
f.write(f"{N} {num_edges}\n")
for v1, v2 in sorted(both_edges):
f.write(f"{v1} {v2}\n")

342
assignments/03/lpa.cpp
View File

@ -1,3 +1,8 @@
#include <algorithm>
#include <array>
#include <cstring>
#include <functional>
#include <limits>
#include <map>
#include <set>
#include <vector>
@ -7,23 +12,15 @@
#include <stdlib.h>
#include <time.h>
#include <unistd.h>
#include <utility>
#ifdef FMT_HEADER_ONLY
#include <fmt/format.h>
#include <fmt/ranges.h>
#endif
// #include <fmt/format.h>
// #include <fmt/ranges.h>
#define TAG_SEND_NUM_EDGES 1001
#define TAG_SEND_EDGES 1002
#define TAG_SEND_FINAL_RESULT 1003
#define MIN(a, b) \
({ \
__typeof__(a) _a = (a); \
__typeof__(b) _b = (b); \
_a < _b ? _a : _b; \
})
typedef struct {
int fst;
int snd;
@ -39,18 +36,9 @@ void pair_vector_init(struct pair_vector *);
void pair_vector_clear(struct pair_vector *);
void pair_vector_push(struct pair_vector *v, int fst, int snd);
pair compute_node_range(int p, int total_num_nodes, int each_num_nodes,
int process);
int lookup_assignment(int *base_node_assignment, pair my_node_range,
std::map<int, std::set<int>> recv_map, int *recvbuf,
int *recv_counts, int *recv_displs, int each_num_nodes,
int rank, int node_number);
int main(int argc, char **argv) {
MPI_Init(&argc, &argv);
int rank, p;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &p);
MPI::Init(argc, argv);
int rank = MPI::COMM_WORLD.Get_rank(), p = MPI::COMM_WORLD.Get_size();
MPI_Datatype IntPairType;
init_pair_type(&IntPairType);
@ -65,15 +53,14 @@ int main(int argc, char **argv) {
pair params;
if (rank == 0) {
printf("Processors: %d, file: %s\n", p, argv[1]);
fp = fopen(argv[1], "r");
// Read the first line
if (getline(&line, &len, fp) != -1)
if ((read = getline(&line, &len, fp)) != -1)
sscanf(line, "%d %d", &params.fst, &params.snd);
}
// Send the params
MPI_Bcast(&params, 1, IntPairType, 0, MPI_COMM_WORLD);
MPI_Bcast(&params, 1, IntPairType, 0, MPI::COMM_WORLD);
int total_num_nodes = params.fst;
int total_num_edges = params.snd;
int each_num_nodes = total_num_nodes / p;
@ -83,15 +70,17 @@ int main(int argc, char **argv) {
rank == p - 1 ? total_num_nodes - rank * each_num_nodes : each_num_nodes;
int my_nodes[num_my_nodes];
pair node_ranges[p];
for (int i = 0; i < p; ++i)
node_ranges[i] = compute_node_range(p, total_num_nodes, each_num_nodes, i);
std::function<std::pair<int, int>(int)> node_range =
[p, total_num_nodes, each_num_nodes](int process) {
int start = process * each_num_nodes;
int end = process == p - 1 ? total_num_nodes : start + each_num_nodes;
return std::make_pair(start, end);
};
// Read the edges
int num_my_edges;
pair *my_edges;
int counts[p], displs[p];
if (rank == 0) {
line = NULL;
// pair all_edges[total_num_edges];
@ -100,31 +89,30 @@ int main(int argc, char **argv) {
// For the current process, what's the last node we're expecting to see?
int current_process = 0;
pair current_node_range = node_ranges[current_process];
std::pair<int, int> current_node_range = node_range(current_process);
int edge_counter = 0;
for (int i = 0; i < total_num_edges; ++i) {
if (getline(&line, &len, fp) == -1)
break;
getline(&line, &len, fp);
int fst, snd;
sscanf(line, "%d %d", &fst, &snd);
if (fst >= current_node_range.snd) {
if (fst >= current_node_range.second) {
if (current_process == 0) {
num_my_edges = edge_counter;
my_edges = (pair *)calloc(num_my_edges, sizeof(pair));
memcpy(my_edges, all_edges.ptr, edge_counter * sizeof(pair));
} else {
MPI_Send(&edge_counter, 1, MPI_INT, current_process,
TAG_SEND_NUM_EDGES, MPI_COMM_WORLD);
TAG_SEND_NUM_EDGES, MPI::COMM_WORLD);
MPI_Send(all_edges.ptr, edge_counter, IntPairType, current_process,
TAG_SEND_EDGES, MPI_COMM_WORLD);
TAG_SEND_EDGES, MPI::COMM_WORLD);
}
// We're starting on the next process
current_process += 1;
current_node_range = node_ranges[current_process];
current_node_range = node_range(current_process);
edge_counter = 0;
pair_vector_clear(&all_edges);
}
@ -141,18 +129,27 @@ int main(int argc, char **argv) {
memcpy(my_edges, all_edges.ptr, edge_counter * sizeof(pair));
} else {
MPI_Send(&edge_counter, 1, MPI_INT, current_process, TAG_SEND_NUM_EDGES,
MPI_COMM_WORLD);
MPI::COMM_WORLD);
MPI_Send(all_edges.ptr, edge_counter, IntPairType, current_process,
TAG_SEND_EDGES, MPI_COMM_WORLD);
TAG_SEND_EDGES, MPI::COMM_WORLD);
}
free(all_edges.ptr);
} else {
MPI_Recv(&num_my_edges, 1, MPI_INT, 0, TAG_SEND_NUM_EDGES, MPI_COMM_WORLD,
MPI_Recv(&num_my_edges, 1, MPI_INT, 0, TAG_SEND_NUM_EDGES, MPI::COMM_WORLD,
NULL);
my_edges = (pair *)calloc(num_my_edges, sizeof(pair));
MPI_Recv(my_edges, num_my_edges, IntPairType, 0, TAG_SEND_EDGES,
MPI_COMM_WORLD, NULL);
MPI::COMM_WORLD, NULL);
}
char *buf = (char *)calloc(sizeof(char), 1000);
int offset = 0; // Keep track of the current position in the buffer
for (int i = 0; i < std::min(num_my_edges, 5); i++) {
offset +=
sprintf(buf + offset, "(%d, %d)", my_edges[i].fst, my_edges[i].snd);
if (i < len - 1) {
// Add a separator (e.g., comma or space) if it's not the last
offset += sprintf(buf + offset, " ");
}
}
if (rank == 0) {
@ -162,25 +159,19 @@ int main(int argc, char **argv) {
}
#pragma endregion
if (rank == 0)
printf("Params: p=%d, |E|=%d, |V|=%d\n", p, total_num_nodes,
total_num_edges);
// STEP 2 TIMER STARTS HERE
MPI_Barrier(MPI_COMM_WORLD);
double step_2_start_time;
if (rank == 0)
step_2_start_time = MPI_Wtime();
MPI::COMM_WORLD.Barrier();
double step_2_start_time = MPI::Wtime();
// Each process analyzes the non-local edges that are contained in its portion
// of the graph.
#pragma region
int node_label_assignment_vec[num_my_nodes];
pair my_node_range = node_ranges[rank];
std::map<int, int> node_label_assignment;
std::pair<int, int> my_node_range = node_range(rank);
// Initial node assignment
for (int idx = 0; idx < num_my_nodes; ++idx) {
node_label_assignment_vec[idx] = my_node_range.fst + idx;
for (int i = my_node_range.first; i < my_node_range.second; ++i) {
node_label_assignment[i] = i;
}
std::map<int, std::set<int>> adj;
@ -191,12 +182,12 @@ int main(int argc, char **argv) {
pair edge = my_edges[i];
adj[edge.fst].insert(edge.snd);
if (!(my_node_range.fst <= edge.fst && edge.fst < my_node_range.snd)) {
if (!(my_node_range.first <= edge.fst && edge.fst < my_node_range.second)) {
non_local_nodes.insert(edge.fst);
non_local_edges.insert(std::make_pair(edge.snd, edge.fst));
}
if (!(my_node_range.fst <= edge.snd && edge.snd < my_node_range.snd)) {
if (!(my_node_range.first <= edge.snd && edge.snd < my_node_range.second)) {
non_local_nodes.insert(edge.snd);
non_local_edges.insert(std::make_pair(edge.fst, edge.snd));
}
@ -212,105 +203,87 @@ int main(int argc, char **argv) {
for (auto entry : non_local_edges) {
int local_node = entry.first, remote_node = entry.second;
int remote_process = remote_node / each_num_nodes;
int corresponding_process = remote_node / each_num_nodes;
// The last process gets some extra nodes
if (remote_process >= p)
remote_process = p - 1;
if (corresponding_process >= p)
corresponding_process = p - 1;
send_map[remote_process].insert(local_node);
recv_map[remote_process].insert(remote_node);
send_map[corresponding_process].insert(local_node);
recv_map[corresponding_process].insert(remote_node);
}
#pragma endregion
// All the processes are communicating to figure out which process needs to
// send what data to the other processes.
#pragma region
// Nothing needs to be done here, I'm using the fact that everything is sent
// in sorted order to ensure that both sides are referring to the same thing
#pragma endregion
// STEP 5 TIMER STARTS HERE
MPI_Barrier(MPI_COMM_WORLD);
double step_5_start_time;
if (rank == 0) {
step_5_start_time = MPI_Wtime();
}
MPI::COMM_WORLD.Barrier();
double step_5_start_time = MPI::Wtime();
// The processes perform the transfers of non-local labels and updates of
// local labels until convergence.
#pragma region
while (true) {
// First, exchange the data that needs to be exchanged
int sendbuf[num_my_nodes];
int send_counts[p];
int send_displs[p];
int recv_counts[p];
int recv_displs[p];
std::map<int, int> remote_labels;
std::vector<int> sendbuf;
std::vector<int> send_counts;
std::vector<int> send_displs;
std::vector<int> recv_counts;
std::vector<int> recv_displs;
if (p > 1) {
int recv_total;
{
int offset = 0;
for (int i = 0; i < p; ++i) {
int count = send_map[i].size();
for (auto local_node : send_map[i]) {
sendbuf[offset + local_node - my_node_range.fst] =
node_label_assignment_vec[local_node - my_node_range.fst];
}
send_counts[i] = count;
send_displs[i] = offset;
offset += count;
int recv_total;
{
int offset = 0;
for (int i = 0; i < p; ++i) {
int count = send_map[i].size();
// std::sort(send_map[i].begin(), send_map[i].end());
for (auto k : send_map[i]) {
sendbuf.push_back(node_label_assignment[k]);
}
offset = 0;
for (int i = 0; i < p; ++i) {
int count = recv_map[i].size();
recv_counts[i] = count;
recv_displs[i] = offset;
offset += count;
}
recv_total = offset;
send_counts.push_back(count);
send_displs.push_back(offset);
offset += count;
}
int recvbuf[recv_total];
MPI_Alltoallv(sendbuf, send_counts, send_displs, MPI_INT, recvbuf,
recv_counts, recv_displs, MPI_INT, MPI_COMM_WORLD);
// Cache efficiently
offset = 0;
for (int i = 0; i < p; ++i) {
std::vector<int> processor_nodes(recv_map[i].begin(),
recv_map[i].end());
for (int j = 0; j < recv_counts[i]; ++j) {
int remote_node = processor_nodes[j];
int remote_value = recvbuf[recv_displs[i] + j];
remote_labels[remote_node] = remote_value;
}
int count = recv_map[i].size();
// std::sort(recv_map[i].begin(), recv_map[i].end());
recv_counts.push_back(count);
recv_displs.push_back(offset);
offset += count;
}
recv_total = offset;
}
std::vector<int> recvbuf(recv_total, 0);
MPI::COMM_WORLD.Alltoallv(sendbuf.data(), send_counts.data(),
send_displs.data(), MPI_INT, recvbuf.data(),
recv_counts.data(), recv_displs.data(), MPI_INT);
std::map<int, int> total_node_label_assignment(node_label_assignment);
for (int i = 0; i < p; ++i) {
std::vector<int> ouais(recv_map[i].begin(), recv_map[i].end());
for (int j = 0; j < recv_counts[i]; ++j) {
int remote_node = ouais[j];
int remote_value = recvbuf[recv_displs[i] + j];
total_node_label_assignment[remote_node] = remote_value;
}
}
// For each local node, determine the minimum label out of its neighbors
std::map<int, int> new_labels;
for (int i = 0; i < num_my_nodes; ++i) {
int node = my_node_range.fst + i;
// int current_value = total_node_label_assignment[i];
int current_value = node_label_assignment_vec[i];
for (int i = my_node_range.first; i < my_node_range.second; ++i) {
int current_value = total_node_label_assignment[i];
int min = current_value;
for (auto neighbor : adj[node]) {
int neighbor_value;
if (my_node_range.fst <= neighbor && neighbor < my_node_range.snd) {
neighbor_value =
node_label_assignment_vec[neighbor - my_node_range.fst];
} else {
neighbor_value = remote_labels[neighbor];
}
// = lookup_assignment(
// node_label_assignment_vec, my_node_range, recv_map,
// recvbuf.data(), recv_counts.data(), recv_displs.data(),
// each_num_nodes, rank, neighbor);
min = MIN(min, neighbor_value);
for (auto neighbor : adj[i]) {
if (total_node_label_assignment[neighbor] < min)
min = total_node_label_assignment[neighbor];
}
if (min < current_value) {
@ -321,8 +294,8 @@ int main(int argc, char **argv) {
// Have there been any changes in the labels?
int num_changes = new_labels.size();
int total_changes;
MPI_Allreduce(&num_changes, &total_changes, 1, MPI_INT, MPI_SUM,
MPI_COMM_WORLD);
MPI::COMM_WORLD.Allreduce(&num_changes, &total_changes, 1, MPI_INT,
MPI::SUM);
if (total_changes == 0) {
break;
@ -330,19 +303,14 @@ int main(int argc, char **argv) {
// Update the original node assignment
for (auto entry : new_labels) {
node_label_assignment_vec[entry.first] = entry.second;
node_label_assignment[entry.first] = entry.second;
}
if (rank == 0)
printf("total changes: %d\n", total_changes);
}
#pragma endregion
// END TIMERS
MPI_Barrier(MPI_COMM_WORLD);
double end_time;
if (rank == 0)
end_time = MPI_Wtime();
MPI::COMM_WORLD.Barrier();
double end_time = MPI::Wtime();
if (rank == 0) {
printf("2-5 Time: %0.04fs\n", end_time - step_2_start_time);
@ -353,36 +321,39 @@ int main(int argc, char **argv) {
// disk.
#pragma region
if (rank == 0) {
FILE *fp = fopen(argv[2], "w");
std::map<int, int> label_count;
for (int process_idx = 0; process_idx < p; ++process_idx) {
pair this_node_range = node_ranges[process_idx];
int count = this_node_range.snd - this_node_range.fst;
if (process_idx == 0) {
std::vector<int> all_assignments(total_num_nodes);
// std::map<int, int> label_count;
int ctr = 0;
for (int i = 0; i < p; ++i) {
std::pair<int, int> this_node_range = node_range(i);
int count = this_node_range.second - this_node_range.first;
if (i == 0) {
for (int j = 0; j < count; ++j) {
fprintf(fp, "%d\n", node_label_assignment_vec[j]);
label_count[node_label_assignment_vec[j]]++;
all_assignments[this_node_range.first + j] =
node_label_assignment[this_node_range.first + j];
// label_count[all_assignments[this_node_range.first + j]]++;
}
} else {
int recvbuf[count];
MPI_Recv(&recvbuf, count, MPI_INT, process_idx, TAG_SEND_FINAL_RESULT,
MPI_COMM_WORLD, NULL);
for (int j = 0; j < count; ++j) {
fprintf(fp, "%d\n", recvbuf[j]);
label_count[recvbuf[j]]++;
}
MPI::COMM_WORLD.Recv(&all_assignments[this_node_range.first], count,
MPI::INT, i, TAG_SEND_FINAL_RESULT);
// for (int j = 0; j < count; ++j) {
// label_count[all_assignments[this_node_range.first + j]]++;
// }
}
}
printf("%d\n", label_count.size());
// std::cout << "Done! " << label_count.size() << std::endl;
} else {
MPI_Send(node_label_assignment_vec, num_my_nodes, MPI_INT, 0,
TAG_SEND_FINAL_RESULT, MPI_COMM_WORLD);
std::vector<int> flat_assignments;
for (int i = my_node_range.first; i < my_node_range.second; ++i) {
flat_assignments.push_back(node_label_assignment[i]);
}
MPI::COMM_WORLD.Send(flat_assignments.data(), flat_assignments.size(),
MPI::INT, 0, TAG_SEND_FINAL_RESULT);
}
#pragma endregion
MPI_Finalize();
MPI::Finalize();
return 0;
}
@ -422,59 +393,4 @@ void pair_vector_push(struct pair_vector *v, int fst, int snd) {
v->ptr[v->len].fst = fst;
v->ptr[v->len].snd = snd;
v->len++;
}
pair compute_node_range(int p, int total_num_nodes, int each_num_nodes,
int process) {
int start = process * each_num_nodes;
int end = process == p - 1 ? total_num_nodes : start + each_num_nodes;
return {.fst = start, .snd = end};
}
int lookup_assignment(int *base_node_assignment, pair my_node_range,
std::map<int, std::set<int>> recv_map, int *recvbuf,
int *recv_counts, int *recv_displs, int each_num_nodes,
int rank, int node_number) {
int process_from = node_number / each_num_nodes;
// Just return from local if local
if (process_from == rank)
return base_node_assignment[node_number - my_node_range.fst];
int count = recv_counts[process_from];
int displs = recv_displs[process_from];
// Determine what index this node is
int index = -1, ctr = 0;
std::vector<int> inner(recv_map[process_from].begin(),
recv_map[process_from].end());
{
// Use binary search...
int lo = 0, hi = count;
while (lo < hi) {
int mid = (lo + hi) / 2;
int midk = inner[mid];
if (node_number < midk)
hi = mid;
else if (node_number > midk)
lo = mid;
else {
index = mid;
break;
}
}
}
// for (int i = 0; i < count; ++i) {
// int remote_node = inner[i];
// if (node_number == remote_node) {
// index = ctr;
// break;
// }
// ctr++;
// }
// Pull the corresponding value from the map
return recvbuf[recv_displs[process_from] + index];
}
}

33
assignments/03/process.py Normal file
View File

@ -0,0 +1,33 @@
import re
WTF = re.compile(r".*: (\d+),.*dataset/(\d+).txt")
by_size = dict()
with open("stdout.txt") as f:
while True:
line1 = f.readline().strip()
if not line1: break
m = WTF.match(line1)
processors = int(m.group(1))
size = int(m.group(2))
if size not in by_size: by_size[size] = dict()
line2 = f.readline().strip()
line3 = f.readline().strip()
time2 = line2.split(": ")[1]
time5 = line3.split(": ")[1]
if processors not in by_size[size]: by_size[size][processors] = (time2, time5)
print("#table(")
print(" columns: (auto, auto, auto, auto, auto, auto),")
columns = [1, 2, 4, 8, 16]
print(" [], ", ", ".join(map(lambda c: f"[{c}]", columns)), ",")
for size, entries in sorted(by_size.items()):
print(f" [{size}],")
for processors, (time2, time5) in sorted(entries.items()):
print(f" [{time2} #linebreak() {time5}],", end = None)
print()
print(")")

66
assignments/03/report.typ
View File

@ -13,5 +13,67 @@ I exchanged data using the unstructured communication approach, doing an
all-to-all transfer.
To read the result efficiently, I tried using the approach given in the slides.
However, this was taking a long time (up to 45 seconds for the 10,000 case) and
I tried using STL's `std::map`. This proved to be orders of magnitude faster
I also tried to use binary search since this would yield $log(n)$ time.
However, this was taking a long time (up to 45 seconds for the 10,000 case), and
it was the bottleneck. Using STL's `std::map` proved to be orders of magnitude
faster.
== Other remarks
On the original example dataset, it poorly using larger numbers. I have an
explanation for this after looking at the performance characteristics of the
run: it completes in one iteration where every single edge is assigned. The data
distribution also indicates that almost everything is connected into the first
node, which isn't balanced.
I've written a generation script in Python using the `igraph` library.
- 1,000: 93 components
- 10,000: 947 components
- 100,000: 9,423 components
- 1,000,000: 92,880 components
Using this data, I was able to achieve much better speedup. I didn't attach the
actual data files but they can be generated with the same script (seeded for
reproducibility).
*NOTE:* I noticed that afterwards, the data was changed again, with a more balanced graph this time.
So the numbers will not reflect the poorer performance.
== Timing on example dataset
This experiment was performed on CSELabs by using my bench script, and the table
was generated with another script.
#table(
columns: (auto, auto, auto, auto, auto, auto),
[], [1], [2], [4], [8], [16] ,
[1000],
[0.0249s #linebreak() 0.0151s],
[0.0234s #linebreak() 0.0122s],
[0.0206s #linebreak() 0.0099s],
[0.0491s #linebreak() 0.0248s],
[0.0177s #linebreak() 0.0106s],
[10000],
[0.2929s #linebreak() 0.1830s],
[0.2933s #linebreak() 0.1540s],
[0.2457s #linebreak() 0.1178s],
[0.3793s #linebreak() 0.1328s],
[0.2473s #linebreak() 0.1197s],
[100000],
[3.7888s #linebreak() 2.4881s],
[3.7592s #linebreak() 2.0212s],
[3.3819s #linebreak() 1.6036s],
[2.9485s #linebreak() 1.3954s],
[2.8593s #linebreak() 1.3107s],
[1000000],
[46.7895s #linebreak() 31.9648s],
[45.2284s #linebreak() 24.8540s],
[40.3994s #linebreak() 20.2851s],
[36.9628s #linebreak() 17.6794s],
[35.7110s #linebreak() 16.6276s],
)