Citation Context

...nd OpenCL [4]. Because of GPU’s high computational power, how to accelerate various workloads on GPUs has been a major research topic in both the high performance computing area and the database area =-=[12, 18, 20, 28, 17, 29, 16, 30, 21, 19, 23, 10, 33, 25, 32]-=-. In high performance computing area, GPUs as accelerators have already been widely deployed to process performance-critical tasks. For example, according to the June 2013’s Top 500 list, more than 50...

Citation Context

...y and flexibility while making the parallelism of the computation explicit to the runtimes. A large number of research efforts have the same goal as Dandelion: building scalable heterogeneous systems =-=[58, 49, 98]-=-. Most of these systems retrofit accelerators such as GPUs, FPGAs, or vector processors into existing frameworks such as MapReduce or MPI. With Dandelion, by contrast, we take on the task of building ...

Citation Context

..., shading, and texture mapping. Second, the computing APIs provide functions for non-graphics applications, such as financial, medical, or weather data analyses [6], database query optimizations [7], =-=[8]-=-, packet routing [9], intrusion detection systems [10], [11], and cryptographic engines [12]– [17]. The most significant differences between the graphics APIs and the computing APIs are sharing and me...

Citation Context

...p. Benchmarks that use relatively high number of dynamic aggregated groups such as bht and regx are more sensitive to AGT size. 6. Related Work Characterization of various benchmarks show that implementations of irregular GPU applications mainly suffer from workload imbalance and scattered memory accesses that cause control flow and memory irregularity [7][9]. Researchers have been investigating and seeking more efficient solutions for these irregular applications by redesigning data structures and re-organizing memory accesses through algorithm, compiler and runtime optimizations [33][40][23][38] to harness the GPU capability. Increasing performance of irregular applications by handling nested parallelism is an important and challenging question in the GPGPU programming community. Lars et al. [6] implement the NESL language on GPU and Lee et al. [20] propose an auto-tuning framework that efficiently maps nested patterns in GPU applications. Yang et al. [39] propose a compiler technique that can activate or disable threads in runtime to handle nested parallelism in GPU applications. Compared with their work, our methodology is to generate workloads dynamically for any nested parallel p...

Citation Context

... overcome this shortcoming by pre-compiling operators. Thus, they only need to compile the query plan itself to a driver program [62]. A similar approach called kernel weaver is used by Wu and others =-=[61]-=-. They combine CUDA kernels for relational primitives into one kernel. This has the advantage that the optimization scope is larger and the compiler can perform more optimizations. However, the disadv...

Citation Context

...End Language Front-End Query Plan Relational Algebra (RA) Primitives RA-to-PTX Kernel Weaver Translation Layer Kernel Weaver: Automatically Perform Kernel Fusion Optimization to reduce data movements =-=[4,5]-=- Oncilla Runtime and API Example Normalized Time �Runtime handles allocation of remote memory and keeps track of connection state for different network layers. �Oncilla API has a concept of “opaque” a...

Citation Context

...rformance improvement over conventional CPUbased solutions. Other studies exploit various software optimization techniques to accelerate query plan generations[32], improve kernel execution efficiency=-=[8;13]-=-, reduce PCIe data transferring[13;33], and support query co-processing with both GPUs and CPUs[34]. Our work in this paper is mainly related to recent development efforts of GPU query engines. Ocelot...

Citation Context

... which combines GPU kernels of relational operators. Their goal is to reduce the data volume that needs to be transfered over the bus and to exploit code optimizations enabled by the combined kernels =-=[39]-=-. Both approaches perform all processing on the GPU and hence, omit possible performance gains due to inter-device parallelism. Przymus and others proposed a bi-objective query planner based on marked...

Citation Context

...formance improvement over conventional CPU-based solutions. Other works exploit various software optimization techniques to accelerate query plan generations [10], improve kernel execution efficiency =-=[30, 26]-=-, reduce PCIe data transferring [30, 31], and support query co-processing with both GPUs and CPUs [22]. Our work in this paper is mainly related to recent development efforts of GPU query engines, whi...