Performance is the most critical aspect towards achieving high scalability of Semantic Web reasoning applications, and considerably limits the application areas of them. There is still a deep mismatch between the requirements for reasoning on a Web scale and performance of the existing reasoning engines. The performance limitation can be considerably reduced by utilizing such large-scale e-Infrastructures as LarKC - the Large Knowledge Collider - an experimental platform for massive distributed incomplete reasoning, which offers several innovative approaches removing the scalability barriers, in particularly, by enabling transparent access to HPC systems. Efficient utilization of such resources is facilitated by means of parallelization being the major element for accomplishing performance and scalability of semantic applications. Here we discuss application of some emerging parallelization strategies and show the benefits obtained by using such systems as LarKC.
infrastructures for large-scale semantic reasoning, such as one set up by LarKC (the
data in heterogeneous data sets. Along with a number of original solutions for obtaining Web scale by semantic applications, LarKC offers services for transparently accessing diverse computing architectures, including multi-core (many-core) multiprocessor, and cluster-based computer architectures as well as dedicated highperformance computers.
operations and thus resolves the conflicts occurring between the concurrent operations while performing computation. Given the large problem sizes that are addressed by LarKC, and considering the benefits of parallelization, it seems natural to explore use of the main parallelization strategies for semantic applications, too. Here we discuss some major parallelization techniques for providing parallelism on task-, instruction-, and data-level, applied for LarKC’s pilot applications. However, the investigated approaches and techniques are quite generic and can be potentially applied for any other Semantic Web engine.
wide range of optimization tasks and might be beneficial for semantic applications. They can be roughly classified according to the level at which the parallelism takes place (Fig. 1): 1) between loosely-coupled components (workflow level) – implementation of parallelism by running multiple instances of the same plug-in simultaneously o task-level parallelism
• workflow branching 2) within a separate component (“plug-in” level) – implementation of parallelism in the concurrent regions of the component’s algorithms o instruction-level parallelism • shared-memory systems: multi-threading • distributed-memory systems: message-passing o data and instruction-level parallelism • MapReduce data processing
Query Transformer
Identifier
Decider Selecter 1 Selecter 2
multi-threading MPI MapReduce
Reasoner cssoe1 r P csseo2 r P ssc3eo r P cssoe4 r P
and obtained performance impact. In this section we discuss only the approaches, which allow obtaining considerable performance impact with a minimum of implementation efforts for sequential code. In particular, we consider multi-threading and message-passing. The achieved performance impact is discussed as well. 3.1 Multi-threading
applications are still using only one of them for their processing (i.e., applications are still sequentially programmed) instead of distributing particular tasks to different processor cores concurrently. In order to make use of the capabilities provided by modern CPU architectures, applications must align their tasks according to the number of available cores.
and does not require much development efforts. For evaluation purposes, we implemented multi-threading support for the Urban Computing application of LarKC [2]. Realization of multi-threading for the most time consuming component of the investigated workflow allowed us to obtain a considerable performance speed-up (Table 1).
The Message-Passing Interface (MPI) is the most widely used parallel programming paradigm for highly-scalable parallel applications. MPI enables sharing the application workload over various nodes of a parallel system (both shared and distributed memory architectures are supported). The synchronization between the nodes is achieved by means of the messages passed among the involved processes through the network interconnect. Implementations of MPI in Java (such as MPIJava or MPJ-Express) have enabled use of MPI also for Java applications. MPI is highly beneficial for computing-intensive applications, whereby scalability within a sharedmemory space is not sufficient for obtaining the necessary performance.
library from the S-Space package1. The parallelization technique was evaluated for the Linked Life Data subset used by University of Sheffield within the LarKC project.
Number of computing nodes 1 2 4 8
750 1 - - -
57 1 20 2.85 10 5.7 5 11.4
strategies, namely multi-threading and message-passing, on performance of two typical Semantic Web use cases. The first application was taken from the urban computing use case, where parallelization facilitates meeting real-time requirements.
performance requirements, applying multi-threading allowed the application to greatly benefit from the multi-core CPU architecture. The second application random indexing - was much more complex as the first one, and made great benefit of message-passing that leveraged a cluster of shared-memory nodes for the application.