Competition GraphHPC-2016 for the fastest implementation of the parallel algorithm Community Detection: Results

In the framework of the GraphHPC-2016 conference held on March 3, 2016 at Moscow State University. Mv Lomonosov at the faculty of the CMC , a competition was held for the fastest implementation of the Community Detection task - the search for communities in an undirected graph with weights.

In the Task section, you can read the problem statement and download an example implementation in C ++. The competition has been held for the second year using an automatic system, the work with which is described in the Information section. This section also provides information about the computing resources of the system, installed software, as well as the requirements for the execution of the task.

The winner of the student nomination received a 512 GB SSD from Samsung. The rest of the winners and prize-winners of the "professional" nomination received Asus ZenWatch smart watches, as well as souvenirs from NVidia and Intel. Also, each of the participants received a "Certificate of Participation."

The results table for one compute node and cluster is as follows:

Computing node:

CPU - 2x Intel Xeon E5-2683 v3 @ 2 GHz, (64 GB RAM, two sockets)

GPU - NVIDIA Tesla K20x (6 GB GDDR5)

No	Participant	Organization	Student	System	Average time, with	Average modularity	Average rating	Algorithm
one	Zhukov Vladimir	VMK MSU	Yes	CPU	0.550	91.10%	12808.14	Label Propagation
2	Kolganov Alexander	VMK MSU		CPU + GPU	0.818	91.23%	11348.78	Louvain
3	Zhuravlev Andrey	MAI	Yes	CPU	0.720	95.79%	9671.70	Label Propagation
four	Alexander Popov	Positive Technologies		CPU	1.657	92.56%	3838.55	Label Propagation
five	Afanasyev Ilya	MSU	Yes	CPU	2.997	90.16%	2066.51	Label Propagation
6	Lamonov Ivan	MSU	Yes	CPU	0.850	24.56%	542.96	Label Propagation
7	Implementation example	DISLab		CPU	3042.287	100.00%	4.75	Louvain

Cluster:

It consists of 36 computing nodes in a 24 + 12 configuration with a high-speed Angara network - the first Russian communication network with VLSI-based routers, where

24 nodes - 2x Intel Xeon CPU E5-2630 @ 2.30GHz (64GB RAM, two sockets)

12 nodes - Intel Xeon CPU E5-2660 @ 2.20GHz (64GB RAM)

No	Participant	Organization	Student	System	Average time, with	Average modularity	Average rating	Algorithm
one	Osipov Artem	T-Platforms		Angara-K1 (32 nodes)	7.668	99.01%	1369.57	Label Propagation
2	Implementation example	DISLab		Angara-K1 (32 nodes)	992.057	79.87%	19.64