high performance computing on graphics processing units: hgpu.org

Multi-GPU systems are widely used in High Performance Computing environments to accelerate scientific computations. This trend is expected to continue as integrated GPUs will be introduced to processors used in multi-socket servers and servers will pack a higher number of GPUs per node. GPUs are currently connected to the system through the PCI Express interconnect, which provides limited bandwidth (compared to the bandwidth of the memory in GPUs) and it often becomes a bottleneck for performance scalability. Current programming models present GPUs as isolated devices with their own memory, even if they share the host memory with the CPU. Programmers explicitly manage allocations in all GPU memories and use primitives to communicate data between GPUs. Furthermore, programmers are required to use mechanisms such as command queues and inter-GPU synchronization. This explicit model harms the maintainability of the code and introduces new sources for potential errors. The first proposal of this thesis is the HPE model. HPE builds a simple, consistent programming interface based on three major features. (1) All device address spaces are combined with the host address space to form a Unified Virtual Address Space. (2) Programs are provided with an Asymmetric Distributed Shared Memory system for all the GPUs in the system. It allows to allocate memory objects that can be accessed by any GPU or CPU. (3) Every CPU thread can request a data exchange between any two GPUs, through simple memory copy calls. Such a simple interface allows HPE to provide always the optimal implementation; eliminating the need for application code to handle different system topologies. Experimental results show improvements on real applications that range from 5% in compute-bound benchmarks to 2.6x in communication-bound benchmarks. HPE transparently implements sophisticated communication schemes that can deliver up to a 2.9x speedup in I/O device transfers. The second proposal of this thesis is a shared memory programming model that exploits the new GPU capabilities for remote memory accesses to remove the need for explicit communication between GPUs. This model turns a multi-GPU system into a shared memory system with NUMA characteristics. In order to validate the viability of the model we also perform an exhaustive performance analysis of remote memory accesses over PCIe. We show that the unique characteristics of the GPU execution model and memory hierarchy help to hide the costs of remote memory accesses. Results show that PCI Express 3.0 is able to hide the costs of up to a 10% of remote memory accesses depending on the access pattern, while caching of remote memory accesses can have a large performance impact on kernel performance. Finally, we introduce AMGE, a programming interface, compiler support and runtime system that automatically executes computations that are programmed for a single GPU across all the GPUs in the system. The programming interface provides a data type for multidimensional arrays that allows for robust, transparent distribution of arrays across all GPU memories. The compiler extracts the dimensionality information from the type of each array, and is able to determine the access pattern in each dimension of the array. The runtime system uses the compiler-provided information to automatically choose the best computation and data distribution configuration to minimize inter-GPU communication and memory footprint. This model effectively frees programmers from the task of decomposing and distributing computation and data to exploit several GPUs. AMGE achieves almost linear speedups for a wide range of dense computation benchmarks on a real 4-GPU system with an interconnect with moderate bandwidth. We show that irregular computations can also benefit from AMGE, too.