Zhenning Wang

Simultaneous Multikernel GPU: Multi-tasking throughput processors via fine-grained sharing

Studies show that non-graphics programs can be less optimized for the GPU hardware, leading to significant resource under-utilization. Sharing the GPU among multiple programs can effectively improve utilization, which is particularly attractive to systems where many applications require access to the GPU (e.g., cloud computing). However, current GPUs lack proper architecture features to support sharing. Initial attempts are preliminary: They either provide only static sharing, which requires recompilation or code transformation, or they do not effectively improve GPU resource utilization. We propose Simultaneous Multikernel (SMK), a fine-grain dynamic sharing mechanism, that fully utilizes resources within a streaming multiprocessor by exploiting heterogeneity of different kernels. We propose several resource allocation strategies to improve system throughput while maintaining fairness. Our evaluation shows that for shared workloads with complementary resource occupancy, SMK improves GPU throughput by 52% over non-shared execution and 17% over a state-of-the-art design.

Simultaneous Multikernel: Fine-Grained Sharing of GPUs

Studies show that non-graphics programs can be less optimized for the GPU hardware, leading to significant resource under-utilization. Sharing the GPU among multiple programs can effectively improve utilization, which is particularly attractive to systems (e.g., cloud computing) where many applications require access to the GPU. However, current GPUs lack proper architecture features to support sharing. Initial attempts are very preliminary in that they either provide only static sharing, which requires recompilation or code transformation, or they do not effectively improve GPU resource utilization. We propose Simultaneous Multikernel (SMK), a fine-grained dynamic sharing mechanism, that fully utilizes resources within a streaming multiprocessor by exploiting heterogeneity of different kernels. We extend the GPU hardware to support SMK, and propose several resource allocation strategies to improve system throughput while maintaining fairness. Our evaluation of 45 shared workloads shows that SMK improves GPU throughput by 34 percent over non-shared execution and 10 percent over a state-of-the-art design.

CAP: co-scheduling based on asymptotic profiling in CPU+GPU hybrid systems

Hybrid systems with CPU and GPU have become the new standard in high performance computing. Workloads are split into two parts and distributed to different devices to utilize both CPU and GPU for data parallelism in hybrid systems. But it is challenging for users to manually balance workload between CPU and GPU since GPU is sensitive to the scale of the problem. Therefore, current dynamic schedulers balance workload between CPU and GPU periodically and dynamically. The periodical balance operation causes frequent synchronizations between CPU and GPU and the synchronizations often degrade the overall performance. To solve the problem, we propose a Co-Scheduling Strategy Based on Asymptotic Profiling (CAP). CAP dynamically splits one tasks workload to CPU and GPU and adopts the profiling technique to predict the workload in next partition. CAP is optimized for GPUs performance characteristics to balance workload between CPU and GPU with only a few synchronizations. We examine our proof-of-concept system with four benchmarks and results show that CAP produces up to 45.1% performance improvement compared with the state-of-art co-scheduling strategy.

CPU+GPU scheduling with asymptotic profiling

Hybrid systems with CPU and GPU have become new standard in high performance computing. Workload can be split and distributed to CPU and GPU to utilize them for data-parallelism in hybrid systems. But it is challenging to manually split and distribute the workload between CPU and GPU since the performance of GPU is sensitive to the workload it received. Therefore, current dynamic schedulers balance workload between CPU and GPU periodically and dynamically. The periodical balance operation causes frequent synchronizations between CPU and GPU. It often degrades the overall performance because of the overhead of synchronizations. To solve the problem, we propose a Co-Scheduling Strategy Based on Asymptotic Profiling (CAP). CAP dynamically splits and distributes the workload to CPU and GPU with only a few synchronizations. It adopts the profiling technique to predict performance and partitions the workload according to the performance. It is also optimized for GPUs performance characteristics. We examine our proof-of-concept system with six benchmarks and evaluation result shows that CAP produces up to 42.7% performance improvement on average compared with the state-of-the-art co-scheduling strategies.

Quality of Service Support for Fine-Grained Sharing on GPUs

GPUs have been widely adopted in data centers to provide acceleration services to many applications. Sharing a GPU is increasingly important for better processing throughput and energy efficiency. However, quality of service (QoS) among concurrent applications is minimally supported. Previous efforts are too coarse-grained and not scalable with increasing QoS requirements. We propose QoS mechanisms for a fine-grained form of GPU sharing. Our QoS support can provide control over the progress of kernels on a per cycle basis and the amount of thread-level parallelism of each kernel. Due to accurate resource management, our QoS support has significantly better scalability compared with previous best efforts. Evaluations show that, when the GPU is shared by three kernels, two of which have QoS goals, the proposed techniques achieve QoS goals 43.8% more often than previous techniques and have 20.5% higher throughput.

DCF: A Dataflow-Based Collaborative Filtering Training Algorithm

Emerging recommender systems often adopt collaborative filtering techniques to improve the recommending accuracy. Existing collaborative filtering techniques are implemented with either alternating least square algorithm or gradient descent (GD) algorithm. However, both of the two algorithms are not scalable because ALS suffers from high computation complexity and GD suffers from severe synchronization problem and tremendous data movement. To solve the above problems, we proposed a Dataflow-based Collaborative Filtering (DCF) algorithm. More specifically, DCF exploits fine-grain asynchronous feature of dataflow model to minimize synchronization overhead; leverages mini-batch technique to reduce computation and communication complexities; uses dummy edge and multicasting techniques to avoid fine-grain overhead of dependency checking and reduce data movement. By utilizing all the above techniques, DCF is able to significantly improve the performance of collaborative filtering. Our experiment on a cluster with one master node and ten slave nodes show that DCF achieves 23

314 Kampferol Inhibits Growth and Invasion of Prostate Cancer via Androgen Dependent Pathway

\times

UP-3.118: Evidence That Estrogens Directly Alter Androgen Regulated Genital Tubercle

× speedup over ALS on Spark and 18