The success of CNNs in various applications is accompanied by a significant increase in the computation and parameter storage costs. Recent efforts toward reducing these overheads involve pruning and compressing the weights of various layers without hurting original accuracy. However, magnitude-based pruning of weights reduces a significant number of parameters from the fully connected layers and may not adequately reduce the computation costs in the convolutional layers due to irregular sparsity in the pruned networks. We present an acceleration method for CNNs, where we prune filters from CNNs that are identified as having a small effect on the output accuracy. By removing whole filters in the network together with their connecting feature maps, the computation costs are reduced significantly. In contrast to pruning weights, this approach does not result in sparse connectivity patterns. Hence, it does not need the support of sparse convolution libraries and can work with existing efficient BLAS libraries for dense matrix multiplications. We show that even simple filter pruning techniques can reduce inference costs for VGG-16 by up to 34% and ResNet-110 by up to 38% on CIFAR10 while regaining close to the original accuracy by retraining the networks.

Human actions often involve complex interactions across several inter-related objects in the scene. However, existing approaches to fine-grained video understanding or visual relationship detection often rely on single object representation or pairwise object relationships. Furthermore, learning interactions across multiple objects in hundreds of frames for video is computationally infeasible and performance may suffer since a large combinatorial space has to be modeled. In this paper, we propose to efficiently learn higher-order interactions between arbitrary subgroups of objects for fine-grained video understanding. We demonstrate that modeling object interactions significantly improves accuracy for both action recognition and video captioning, while saving more than 3-times the computation over traditional pairwise relationships. The proposed method is validated on two large-scale datasets: Kinetics and ActivityNet Captions. Our SINet and SINet-Caption achieve state-of-the-art performances on both datasets even though the videos are sampled at a maximum of 1 FPS. To the best of our knowledge, this is the first work modeling object interactions on open domain large-scale video datasets, and we additionally model higher-order object interactions which improves the performance with low computational costs. * Work performed as a NEC Labs intern Attended interactions Interactions between ROIs Video frame Action prediction: cooking on campfire , cooking egg , …

Hardware devices can fail, but many drivers assume they do not. When confronted with real devices that misbehave, these assumptions can lead to driver or system failures. While major operating system and device vendors recommend that drivers detect and recover from hardware failures, we find that there are many drivers that will crash or hang when a device fails. Such bugs cannot easily be detected by regular stress testing because the failures are induced by the device and not the software load.This paper describes Carburizer, a code-manipulation tool and associated runtime that improves system reliability in the presence of faulty devices. Carburizer analyzes driver source code to find locations where the driver incorrectly trusts the hardware to behave. Carburizer identified almost 1000 such bugs in Linux drivers with a false positive rate of less than 8 percent. With the aid of shadow drivers for recovery, Carburizer can automatically repair 840 of these bugs with no programmer involvement.To facilitate proactive management of device failures, Carburizer can also locate existing driver code that detects device failures and inserts missing failure-reporting code. Finally, the Carburizer runtime can detect and tolerate interrupt-related bugs, such as stuck or missing interrupts.

No abstract

Virtual machine migration greatly aids management by allowing flexible provisioning of resources and decommissioning of hardware for maintenance. However, efforts to improve network performance by granting virtual machines direct access to hardware currently prevent migration. This occurs because (1) the VMM cannot migrate the state of the device, and (2) the source and destination machines may have different network devices, requiring different drivers to run in the migrated virtual machine.In this paper, we describe a lightweight software mechanism for migrating virtual machines with direct hardware access. We base our solution on shadow drivers, which efficiently capture and reset the state of a device driver. On the source hardware, the shadow driver continuously monitors the state of the driver and device. After migration, the shadow driver uses this information to configure a driver for the corresponding device on the destination. We implement our solution for Linux network drivers running on the Xen hypervisor. We show that the performance overhead, compared to direct hardware access, is negligible and is much better than using a virtual NIC.

No abstract

No abstract

Device drivers are the single largest contributor to operating-system kernel code with over 5 million lines of code in the Linux kernel, and cause significant complexity, bugs and development costs. Recent years have seen a flurry of research aimed at improving the reliability and simplifying the development of drivers. However, little is known about what constitutes this huge body of code beyond the small set of drivers used for research.In this paper, we study the source code of Linux drivers to understand what drivers actually do, how current research applies to them and what opportunities exist for future research. We determine whether assumptions made by driver research, such as that all drivers belong to a class, are indeed true. We also analyze driver code and abstractions to determine whether drivers can benefit from code re-organization or hardware trends. We develop a set of staticanalysis tools to analyze driver code across various axes. Broadly, our study looks at three aspects of driver code (i) what are the characteristics of driver code functionality and how applicable is driver research to all drivers, (ii) how do drivers interact with the kernel, devices, and buses, and (iii) are there similarities that can be abstracted into libraries to reduce driver size and complexity?We find that many assumptions made by driver research do not apply to all drivers. At least 44% of drivers have code that is not captured by a class definition, 28% of drivers support more than one device per driver, and 15% of drivers do significant computation over data. From the driver interactions study, we find that the USB bus offers an efficient bus interface with significant standardized code and coarse-grained access, ideal for executing drivers in isolation. We also find that drivers for different buses and classes have widely varying levels of device interaction, which indicates that the cost of isolation will vary by class. Finally, from our driver similarity study, we find 8% of all driver code is substantially similar to code elsewhere and may be removed with new abstractions or libraries.