Network Function Virtualization (NFV) paradigm offers flexibility, cost benefits, and ease of deployment by decoupling network function from hardware middleboxes. The service function chains (SFC) deployed using the NFV platform require efficient sharing of resources among various network functions in the chain. Graphics Processing Units (GPUs) have been used to improve various network functions’ performance. However, sharing a single GPU among multiple virtualized network functions (virtual machines) in a service function chain has been challenging due to their proprietary hardware and software stack. Earlier GPU architectures had a limitation: a single physical GPU can only be allocated to one virtual machine (VM) and cannot be shared among multiple VMs. The newer GPUs are virtualization-aware (hardware-assisted virtualization) and allow multiple virtual machines to share a single physical GPU. Although virtualization-aware, these GPUs still lack support for custom scheduling policies and do not expose the preemption control to users. When network functions (hosted within virtual machines) with different processing requirements share the same GPU, virtualization-aware GPUs’ default round-robin scheduling mechanism proves to be inefficient, resulting in packet drops and lower throughput. This paper presents Simmer, an efficient mechanism for scheduling a network function service chain on virtualization-aware GPUs. Our scheduling solution considers the processing requirement of NFs in a GPU-based SFC, thus improving overall throughput by up to 29% and reducing the packet drop to zero compared to vanilla setup.

As the popularity and relevance of the Function-as-a-Service (FaaS) model keeps growing, we believe newer avatars of the service will support computationally intensive SIMT functions that will execute on GPUs. With hardware-assisted virtualization of GPUs now possible, cloud offerings including GPUs usually bind a virtual GPU (vGPU) to a VM. While there is a choice of scheduling algorithms to multiplex vGPUs on to the physical GPU, the work-conserving best-effort scheduler helps to maintain a high level of utilization of the GPU. With this, we observe that the total share of the GPU per VM is non-deterministic and depends on how different VMs load the GPU via their vGPUs. As a result, any function-to-vGPU scheduler that does not explicitly account for this nondeterministic vGPU capacity will suffer from lower than optimal goodput - particularly when these functions are deadline bound as is the case with FaaS offerings today. In this work, we exploit a software based task slicing technique to dynamically determine task sizes for scheduling on vGPUs to maximize successful completion of functions within their deadlines. Our solution extends the conventional earliest-deadline first (EDF) scheduling algorithm by balancing scheduling opportunities (via kernel slicing) and maximizing the chances of functions finishing before their deadline. The work is motivated by the fact that static decisions that consider entire tasks as scheduling units or use a fixed, statically decided slice size as scheduling units cannot adapt to the non-deterministic vGPU capacity. A comparison of our solution with well-known deadline aware scheduling approach (earliest deadline first), yielded an improvement of up to 2.9x in goodput.

The increasing use of Graphics Processing Unit (GPUs) for accelerating compute intensive tasks and graphics-related computations has led to their inclusion in High Performance Clusters and Cloud setups. Several cloud vendors provide virtual machine instances with GPU capabilities. With the advent of virtualization aware GPU hardware (NVIDIA vGPUs), allocating and sharing physical GPU among virtual machines has become easier and cost-efficient. The sharing mechanism and extent of sharing are determined by the vGPU scheduling algorithm and a configurable vGPU profile. As part of this work, we present a thorough empirical study of the hardware-assisted virtualized GPU setups. In particular, we quantify the virtualization overheads, study the interference effects of concurrently executing homogeneous and heterogeneous workloads, and impact of vGPU scheduling algorithms. We also demonstrate that the best vGPU configuration parameters are sensitive to the mix of workload characteristics that share the GPU. Our study also compares the performance of vGPU based virtualization with PCI passthrough based direct GPU assignment to virtual machines. Based on our evaluation using heterogeneous workloads setups and varying vGPU configurations, we observe that the virtualization overheads are up to 7% (in terms of reduced memory availability) and up to 20% increase in execution times. Further, we demonstrate using our setups, that co-placing of heterogeneous workloads can improve the efficiency of GPU multiplexing and decrease execution times to as low as 20% as compared to homogeneous workload placements.

Statistical shape models, such as Active Shape Models (ASMs), suffer from their inability to represent a large range of variations of a complex shape and to account for the large errors in detection of model points. We propose a novel method (dubbed PDM-ENLOR) that overcomes these limitations by locating each shape model point individually using an ensemble of local regression models and appearance cues from selected model points. Our method first detects a set of reference points which were selected based on their saliency during training. For each model point, an ensemble of regressors is built. From the locations of the detected reference points, each regressor infers a candidate location for that model point using local geometric constraints, encoded by a point distribution model (PDM). The final location of that point is determined as a weighted linear combination, whose coefficients are learnt from the training data, of candidates proposed from its ensemble's component regressors. We use different subsets of reference points as explanatory variables for the component regressors to provide varying degrees of locality for the models in each ensemble. This helps our ensemble model to capture a larger range of shape variations as compared to a single PDM. We demonstrate the advantages of our method on the challenging problem of segmenting gene expression images of mouse brain.

Studies have shown that aortic calcification is associated with cardiovascular disease. In this study, a method for localization, centerline extraction, and segmentation of the thoracic aorta in noncontrast cardiac-computed tomography (CT) images, toward the detection of aortic calcification, is presented. The localization of the right coronary artery ostium slice is formulated as a regression problem whose input variables are obtained from simple intensity features computed from a pyramid representation of the slice. The localization, centerline extraction, and segmentation of the aorta are formulated as optimal path detection problems. Dynamic programming is applied in the Hough space for localizing key center points in the aorta which guide the centerline tracing using a fast marching-based minimal path extraction framework. The input volume is then resampled into a stack of 2-D cross-sectional planes orthogonal to the obtained centerline. Dynamic programming is again applied for the segmentation of the aorta in each slice of the resampled volume. The obtained segmentation is finally mapped back to its original volume space. The performance of the proposed method was assessed on cardiac noncontrast CT scans and promising results were obtained.

Analysis of gene expression patterns in brain images obtained from high-throughput in situ hybridization requires accurate and consistent annotations of anatomical regions/subregions. Such annotations are obtained by mapping an anatomical atlas onto the gene expression images through intensity- and/or landmark-based registration methods or deformable model-based segmentation methods. Due to the complex appearance of the gene expression images, these approaches require a pre-processing step to determine landmark correspondences in order to incorporate landmark-based geometric constraints. In this paper, we propose a novel method for landmark-constrained, intensity-based registration without determining landmark correspondences a priori. The proposed method performs dense image registration and identifies the landmark correspondences, simultaneously, using a single higher-order Markov Random Field model. In addition, a machine learning technique is used to improve the discriminating properties of local descriptors for landmark matching by projecting them in a Hamming space of lower dimension. We qualitatively show that our method achieves promising results and also compares well, quantitatively, with the expert's annotations, outperforming previous methods.

An accurate labeling of a multi-part, complex anatomical structure (e.g., brain) is required in order to compare data across images for spatial analysis. It can be achieved by fitting an object-specific geometric atlas that is constructed using a partitioned, high-resolution deformable mesh and tagging each of its polygons with a region label. Subdivision meshes have been used to construct such an atlas because they can provide a compact representation of a partitioned, multi-resolution, object-specific mesh structure using only a few control points. However, automated fitting of a subdivision mesh-based geometric atlas to an anatomical structure in an image is a difficult problem and has not been sufficiently addressed. In this paper, we propose a novel Markov Random Field-based method for fitting a planar, multi-part subdivision mesh to anatomical data. The optimal fitting of the atlas is obtained by determining the optimal locations of the control points. We also tackle the problem of landmark matching in tandem with atlas fitting by constructing a single graphical model to impose pose-invariant, landmark-based geometric constraints on atlas deformation. The atlas deformation is also governed by additional constraints imposed by the mesh's geometric properties and the object boundary. We demonstrate the potential of the proposed method on the difficult problem of segmenting a mouse brain and its interior regions in gene expression images which exhibit large intensity and shape variability. We obtain promising results when compared with manual annotations and prior methods.

In this paper, we present a graph-based image segmentation method (patch-cuts) that incorporates features and spatial relations obtained from image patches. In the first step, patch-cuts extracts a set of patches that can assume arbitrary shape and size. Patches are determined by a combination of intensity quantization and morphological operations and render the proposed method robust against noise. Upon patch extraction, a set of intensity, texture and shape features are computed for each patch. These features are integrated and minimized simultaneously in a tunable energy function. Patch-cuts explores the benefit of information theory-based measures such as the Kullback-Leibler and the Jensen-Shannon divergence in its energy terms. In our experiments, we applied patchcuts to general images as well as to non-contrast Computed Tomography heart scans. 1

Image segmentation is, in general, an ill-posed problem and additional constraints need to be imposed in order to achieve the desired result. Particularly in the field of medical image segmentation, a significant amount of prior knowledge is available that can be used to constrain the solution space of the segmentation problem. However, most of this prior knowledge is, in general, vague or imprecise in nature, which makes it very difficult to model. This is the problem that is addressed in this paper. Specifically, in this paper, we present fuzzy-cuts, a novel, knowledge-driven, graph-based method for medical image segmentation. We cast the problem of image segmentation as the maximum a posteriori (MAP) estimation of a Markov random field (MRF) which, in essence, is equivalent to the minimization of the corresponding Gibbs energy function. Considering the inherent imprecision that is common in the a priori description of objects in medical images, we propose a fuzzy theoretic model to incorporate knowledge-driven constraints into the MAP-MRF formulation. In particular, we focus on prior information about the object's location, appearance and spatial connectivity to a known seed region inside the object. To that end, we introduce fuzzy connectivity and fuzzy location priors that are used in combination to define the first-order clique potential of the Gibbs energy function. In our experiments, we demonstrate the application of the proposed method to the challenging problem of heart segmentation in non-contrast computed tomography (CT) data.

In spite of the advancement and proliferation of cardiovascular imaging, the rate of deaths due to unpredicted heart attack remains high. Thus, it becomes imperative to develop novel computational tools to mine quantitative parameters from the imaging data for early detection of asymptomatic cardiovascular disease. Coronary calcification burden has been reported to be a significant and independent predictor of the atherosclerosis disease, and is associated with future cardiac events. Additionally, increased neovascularization of the plaque has been identified as a common feature of coronary plaque inflammation and has been defined as a plaque vulnerability index. In this paper, we present methods to extract and quantify coronary calcifications in the non-contrast cardiac CT scans and to detect neovascularization in the coronary vessels using contrast-enhanced intra-vascular ultrasound imaging.

In this paper we describe a new hardware/software partitioning approach for embedded reconfigurable architectures consisting of a general-purpose processor (CPU), a dynamically reconfigurable datapath (e.g. an FPGA), and a memory hierarchy. We have developed a framework called Nimble that automatically compiles system-level applications specified in C to executables on the target platform. A key component of this framework is a hardware/software partitioning algorithm that performs fine-grained partitioning (at loop and basic-block levels) of an application to execute on the combined CPU and datapath. The partitioning algorithm optimizes the global application execution time, including the software and hardware execution times, communication time and datapath reconfiguration time. Experimental results on real applications show that our algorithm is effective in rapidly finding close to optimal solutions.