STOC: Secure, Tactical On-Demand Cloud

ABSTRACT: We present a novel architecture for on-demand clouds, which means creating a cloud computing environment when needed that opportunistically takes advantage of available processors. The processors are located on mobile computers at the tactical edge of a network and connect via wireless networks. Clusters of processors can be interconnected by trunks. Compute nodes may appear or disappear unpredictably (e.g., nodes may be on disposable handhelds or moving Unmanned Aerial Systems, or may be damaged in a combat environment). Thus we examine survivable fault tolerant parallel computing frameworks, such as MapReduce, which can adapt resources on the fly when processors fail. We focus on exploiting Graphical Processing Units (GPUs), which offer a greater density of processing cores compared to CPUs given the same physical space limits, and are well suited to many numerical calculations (e.g., data fusion) involved with sensor data. We examine how to implement MapReduce for GPUs in a new way that takes advantage of emerging capabilities in GPU instruction sets to avoid multi-pass requirements of past work in this area. We devise algorithms that can seamlessly combine GPUs and CPUs by using the OpenCL language for coding. Our target problem is distributed 3D scene reconstruction on demand. BENEFIT: This work adapts clouds to a tactical edge environment, where commercial clouds (e.g., from Amazon, Google) are not designed to work. One benefit is to enable on-demand cloud computing with virtualization that offers secure processing on an untrusted infrastructure. This includes security controls providing confidentiality and integrity, verified through cryptographic proofs in accordance with NIST 800-53. Through the use of GPU chips, another benefit is enabling real-time response to decentralized tactical users by exploiting more massive parallelism for a given size, weight, and power limit than conventional Central Processing Units (CPUs) can achieve. GPU chips are approaching a thousand or more stream cores (e.g., 6 GFLOPS [giga floating point operations per second] per watt for one chip). We also allow cloud computing with seamless distribution of computation over heterogeneous GPU/CPU nodes, which is ideal for a tactical setting that may combine various types of hardware devices with and without GPUs. We also simplify the MapReduce framework for end users to allow users with lower expertise and programming skills to configure computations; this allows faster deployment of new capabilities on our cloud architecture.