Exploring HPC and Big Data Convergence: A Graph Processing Study on Intel Knights Landing

Abstract

The question "Can big data and HPC infrastructure converge?" has important implications for many operators and clients of modern computing. However, answering it is challenging. The hardware is currently different, and fast evolving: big data uses machines with modest numbers of fat cores per socket, large caches, and much memory, whereas HPC uses machines with larger numbers of (thinner) cores, non-trivial NUMA architectures, and fast interconnects. In this work, we investigate the convergence of big data and HPC infrastructure for one of the most challenging application domains, the highly irregular graph processing. We contrast through a systematic, experimental study of over 300,000 core-hours the performance of a modern multicore, Intel Knights Landing (KNL) and of traditional big data hardware, in processing representative graph workloads using state-of-the-art graph analytics platforms. The experimental results indicate KNL is convergence-ready, performance-wise, but only after extensive and expert-level tuning of software and hardware parameters.

The future is big graphs

“…Their interdependency forms the Hardware-Platform-Algorithm-Dataset (HPAD) Law. 29 3. Graph processing uses a complex pipeline, combining a variety of tasks other than querying and algorithms.…”

A survey on evaluating and optimizing performance of Intel Xeon Phi

“…Overall, Phi does not provide comparable performance to CPU as a stand-alone shared memory processor. Thread-affinity strategy Balanced, 4,14,20,21,23,26,36 scatter, 37 compact, 92 no single winner 13,84,94 Memory mode Cache, 60,62 flat, 55,90,96 hybrid (none), no single winner [10][11][12]52,54,57,97 Interconnect clustering mode All-to-all (none), quadrant, 11,62 sub-NUMA, 10,55,57,96,97 no single winner 52,96…”

“…Also, some works use OpenMP task pragmas for parallelization. 33,52,63 TA B L E 5 Parallelization language/library used on Phi MPI 4,9,12,16,22,24,25,28,35,36,43,44,50,54,59,60,63,66,76,82,86 Others Pthreads, 11,23,76,84,91,95 Intel TBB, 8,48 Cilk Plus, 48 OpenCL 49,100 Chatzikonstantis et al 28 study inferior-olivary nucleus (InfOli) simulation which is used in brain modeling. They accelerate the simulation using (i) MPI, (ii) OpenMP, and (iii) hybrid MPI+OpenMP.…”

“…However, achieving peak performance on Phi is not easy, as it requires myriad set of optimizations and understanding of the microarchitecture. 11,94 The pragma-based programming style makes the code complex, especially for applications that use both CPU and Phi. Large coding-efforts and difficulties in debugging reduce programmer productivity and create a lag between progress in algorithm/application-domain and their acceleration on Phi.…”

Accelerating Spark-Based Applications with MPI and OpenACC

“…Currently, Apache Spark is the largest and perhaps the most comprehensive open-source framework dealing with big data [6]. However, due to its use of the Java Virtual Machine (JVM), its performance is not as desired, and there is still much room for improvement.…”