Addressing Memory Bottlenecks for Emerging Applications

There has been a recent emergence of applications from the domain of machine learning, data mining, numerical analysis and image processing. These applications are becoming the primary algorithms driving many important user-facing applications and becoming pervasive in our daily lives. Due to their increasing usage in both mobile and datacenter workloads, it is necessary to understand the software and hardware demands of these applications, and design techniques to match their growing needs. This dissertation studies the performance bottlenecks that arise when we try to improve the performance of these applications on current hardware systems. We observe that most of these applications are data-intensive, i.e., they operate on a large amount of data. Consequently, these applications put significant pressure on the memory. Interestingly, we notice that this pressure is not just limited to one memory structure. Instead, different applications stress different levels of the memory hierarchy. For example, training Deep Neural Networks (DNN), an emerging machine learning approach, is currently limited by the size of the GPU main memory. On the other spectrum, improving DNN inference on CPUs is bottlenecked by Physical Register File (PRF) bandwidth. Concretely, this dissertation tackles four such memory bottlenecks for these emerging applications across the memory hierarchy (off-chip memory, on-chip memory and physical register file), presenting hardware and software techniques to address these bottlenecks and improve the performance of the emerging applications. For on-chip memory, we present two scenarios where emerging applications perform at a sub-optimal performance. First, many applications have a large number of marginal bits that do not contribute to the application accuracy, wasting unnecessary space and transfer costs. We present ACME, an asymmetric compute-memory paradigm, that removes marginal bits from the memory hierarchy while performing the computation in full precision. Second, we tackle the contention in shared caches for these emerging applications that arise in datacenters where multiple applications can share the same cache capacity. We present ShapeShifter, a runtime system that continuously monitors the runtime environment, detects changes in the cache availability and dynamically recompiles the application on the fly to efficiently utilize the cache capacity. For physical register file, we observe that DNN inference on CPUs is primarily limited by the PRF bandwidth. Increasing the number of compute units in CPU requires increasing the read ports in the PRF. In this case, PRF quickly reaches a point where latency could no longer be met. To solve this problem, we present LEDL, locality extensions for deep learning on CPUs, that entails a rearchitected FMA and PRF design tailored for the heavy data reuse inherent in DNN inference. Finally, a significant challenge facing both the researchers and industry practitioners is that as the DNNs grow deeper and larger, the DNN training is limited by the size of the GPU main memory, restricting the size of the networks which GPUs can train. To tackle this challenge, we first identify the primary contributors to this heavy memory footprint, finding that the feature maps (intermediate layer outputs) are the heaviest contributors in training as opposed to the weights in inference. Then, we present Gist, a runtime system, that uses three efficient data encoding techniques to reduce the footprint of DNN training.PhDComputer Science & EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttps://deepblue.lib.umich.edu/bitstream/2027.42/146016/1/anijain_1.pd