Transactional memory is a concurrency control mechanism that
aims to simplify the development of parallel applications in the
times of ever-increasing core counts of multi-processor architec-
tures. At the same time, it must support the application to achieve a
high level of performance and scalability to justify it as an alterna-
tive to classical lock-based synchronization. These claims still lack
a broad validation in the field because the high number of proposed
transactional memory algorithms are evaluated using only a small
variety of existing benchmark applications.
In this paper, we introduce DREAM, an extensible benchmark
framework for processing live updates to large graphs using trans-
actional memory. It represents real-world workloads from the do-
main of data stream processing. Popular social networks, like Face-
book, Twitter or Google+ have increased the economical interest in
computing on large graphs. However, social networks are scale-free
- i.e. their degree distribution follows a power-law: The majority
of edges connect to the minority of vertices, which makes compu-
tations on these networks hard to scale. On the other hand, “Big
Graphs” don’t necessarily mean “Big Data”. Even huge networks
can be processed efficiently on a single machine, especially when
done in parallel.
We currently support in the DREAM framework two opera-
tors that implement standard graph algorithms which operate on
streamed updates: community of interest and community detec-
tion based on label propagation. However, DREAM is designed
very modular and new algorithms can be added straightforward
by implementing them as new operators. We evaluate the use of
DREAM with well-known software transactional memory imple-
mentations.