Graph processing is a vital component of many AI and big data applications. However, due to its poor locality and complex data access patterns, graph processing is also a known performance killer of AI and big data applications. In this work, we propose to enhance graph processing applications by leveraging fine-grained memory access patterns with a dual-path architecture on top of existing software-based graph optimizations. We first identify that memory accesses to the offset, edge, and state array have distinct locality and impact on performance. We then introduce the Skyway architecture, which consists of two primary components: 1) a dedicated direct data path between the core and memory to transfer state array elements efficiently, and 2) a data-type aware fine-grained memory-side row buffer hardware for both the newly designed direct data path and the regular memory hierarchy data path. The proposed Skyway architecture is able to improve the overall performance by reducing the memory access interference and improving data access efficiency with a minimal overhead. We evaluate Skyway on a set of diverse algorithms using large real-world graphs. On a simulated four-core system, Skyway improves the performance by 23% on average over the best-performing graph-specialized hardware optimizations.

The dataflow architecture, which is characterized by a lack of a redundant unified control logic, has been shown to have an advantage over the control-flow architecture as it improves the computational performance and power efficiency, especially of applications used in high-performance computing (HPC). Importantly, the high computational efficiency of systems using the dataflow architecture is achieved by allowing program kernels to be activated in a simultaneous manner. Therefore, a proper acknowledgment mechanism is required to distinguish the data that logically belongs to different contexts. Possible solutions include the tagged-token matching mechanism in which the data is sent before acknowledgments are received but retried after rejection, or a handshake mechanism in which the data is only sent after acknowledgments are received. However, these mechanisms are characterized by both inefficient data transfer and increased area cost. Good performance of the dataflow architecture depends on the efficiency of data transfer. In order to optimize the efficiency of data transfer in existing dataflow architectures with a minimal increase in area and power cost, we propose a Look-Ahead Acknowledgment (LAA) mechanism. LAA accelerates the execution flow by speculatively acknowledging ahead without penalties. Our simulation analysis based on a handshake mechanism shows that our LAA increases the average utilization of computational units by 23.9%, with a reduction in the average execution time by 17.4% and an increase in the average power efficiency of dataflow processors by 22.4%. Crucially, our novel approach results in a relatively small increase in the area and power consumption of the on-chip logic of less than 0.9%. In conclusion, the evaluation results suggest that Look-Ahead Acknowledgment is an effective improvement for data transfer in existing dataflow architectures.

With the coming of exascale supercomputing era, power efficiency has become the most important obstacle to build an exascale system. Dataflow architecture has native advantage in achieving high power efficiency for scientific applications. However, the state-of-the-art dataflow architectures fail to exploit high parallelism for loop processing. To address this issue, we propose a pipelining loop optimization method (PLO), which makes iterations in loops flow in the processing element (PE) array of dataflow accelerator. This method consists of two techniques, architecture-assisted hardware iteration and instruction-assisted software iteration. In hardware iteration execution model, an on-chip loop controller is designed to generate loop indexes, reducing the complexity of computing kernel and laying a good foundation for pipelining execution. In software iteration execution model, additional loop instructions are presented to solve the iteration dependency problem. Via these two techniques, the average number of instructions ready to execute per cycle is increased to keep floating-point unit busy. Simulation results show that our proposed method outperforms static and dynamic loop execution model in floating-point efficiency by 2.45x and 1.1x on average, respectively, while the hardware cost of these two techniques is acceptable.

Double buffering is an effective mechanism to hide the latency of data transfers between on-chip and off-chip memory. However, in dataflow architecture, the swapping of two buffers during the execution of many tiles decreases the performance because of repetitive filling and draining of the dataflow accelerator. In this work, we propose a non-stop double buffering mechanism for dataflow architecture. The proposed non-stop mechanism assigns tiles to the processing element array without stopping the execution of processing elements through optimizing control logic in dataflow architecture. Moreover, we propose a work-flow program to cooperate with the non-stop double buffering mechanism. After optimizations both on control logic and on work-flow program, the filling and draining of the array needs to be done only once across the execution of all tiles belonging to the same dataflow graph. Experimental results show that the proposed double buffering mechanism for dataflow architecture achieves a 16.2% average efficiency improvement over that without the optimization.