No abstract

Cache partitioning is now available in commercial hardware. In theory, software can leverage cache partitioning to use the last-level cache better and improve performance. In practice, however, current systems implement way-partitioning, which offers a limited number of partitions and often hurts performance. These limitations squander the performance potential of smart cache management.We present KPart, a hybrid cache partitioning-sharing technique that sidesteps the limitations of way-partitioning and unlocks significant performance on current systems. KPart first groups applications into clusters, then partitions the cache among these clusters. To build clusters, KPart relies on a novel technique to estimate the performance loss an application suffers when sharing a partition. KPart automatically chooses the number of clusters, balancing the isolation benefits of waypartitioning with its potential performance impact. KPart uses detailed profiling information to make these decisions. This information can be gathered either offline, or online at low overhead using a novel profiling mechanism.We evaluate KPart in a real system and in simulation. KPart improves throughput by 24% on average (up to 79%) on an Intel Broadwell-D system, whereas prior per-application partitioning policies improve throughput by just 1.7% on average and hurt 30% of workloads. Simulation results show that KPart achieves most of the performance of more advanced partitioning techniques that are not yet available in hardware.

Graph processing is increasingly bottlenecked by main memory accesses. On-chip caches are of little help because the irregular structure of graphs causes seemingly random memory references. However, most real-world graphs offer significant potential locality-it is just hard to predict ahead of time. In practice, graphs have well-connected regions where relatively few vertices share edges with many common neighbors. If these vertices were processed together, graph processing would enjoy significant data reuse. Hence, a graph's traversal schedule largely determines its locality.This paper explores online traversal scheduling strategies that exploit the community structure of real-world graphs to improve locality. Software graph processing frameworks use simple, locality-oblivious scheduling because, on general-purpose cores, the benefits of locality-aware scheduling are outweighed by its overheads. Software frameworks rely on offline preprocessing to improve locality. Unfortunately, preprocessing is so expensive that its costs often negate any benefits from improved locality. Recent graph processing accelerators have inherited this design. Our insight is that this misses an opportunity: Hardware acceleration allows for more sophisticated, online locality-aware scheduling than can be realized in software, letting systems significantly improve locality without any preprocessing.To exploit this insight, we present bounded depth-first scheduling (BDFS), a simple online locality-aware scheduling strategy. BDFS restricts each core to explore one small, connected region of the graph at a time, improving locality on graphs with good community structure. We then present HATS, a hardwareaccelerated traversal scheduler that adds just 0.4% area and 0.2% power over general-purpose cores.We evaluate BDFS and HATS on several algorithms using large real-world graphs. On a simulated 16-core system, BDFS reduces main memory accesses by up to 2.4× and by 30% on average. However, BDFS is too expensive in software and degrades performance by 21% on average. HATS eliminates these overheads, allowing BDFS to improve performance by 83% on average (up to 3.1×) over a locality-oblivious software implementation and by 31% on average (up to 2.1×) over specialized prefetchers.

No abstract

Convolutional Neural Networks (CNNs) have emerged as a fundamental technology for machine learning. High performance and extreme energy efficiency are critical for deployments of CNNs, especially in mobile platforms such as autonomous vehicles, cameras, and electronic personal assistants. This paper introduces the Sparse CNN (SCNN) accelerator architecture, which improves performance and energy efficiency by exploiting the zero-valued weights that stem from network pruning during training and zero-valued activations that arise from the common ReLU operator. Specifically, SCNN employs a novel dataflow that enables maintaining the sparse weights and activations in a compressed encoding, which eliminates unnecessary data transfers and reduces storage requirements. Furthermore, the SCNN dataflow facilitates efficient delivery of those weights and activations to a multiplier array, where they are extensively reused; product accumulation is performed in a novel accumulator array. On contemporary neural networks, SCNN can improve both performance and energy by a factor of 2.7x and 2.3x, respectively, over a comparably provisioned dense CNN accelerator.