Abstract: Certain aspects of the present disclosure provide techniques and apparatus for quantized machine learning. A quantized input matrix is accessed at a layer of a neural network, and a first interim value is generated in an accumulator by performing matrix multiplication, using the accumulator, of the quantized input matrix and a quantized weight matrix associated with the layer of the neural network. The first interim value is normalized based at least in part on one or more leading sign bits of the first interim value, and the normalized first interim value is dequantized. A second interim value is generated by applying a rounded right-shift operation to the dequantized normalized first interim value, and activation data is generated by applying an activation function to the second interim value.

Abstract: A node having a sensor and a computing device is provided for identifying a location of a source of a physical process. The node collects via the sensor a measurement of the physical process. The node repeatedly recalculates parameters until termination criterion is satisfied. The node receives parameters of a probability distribution and a gradient from other nodes. The node generates parameters based on the parameters and the gradients received from the other nodes. The node samples from a distribution of source locations based on the generated parameters. The node calculates a gradient derived from the sampled source locations, the generated parameters, and a joint probability of the sampled source locations and the measurement. The node sends to a subset of other nodes the generated parameters and the calculated gradient. When the termination criterion is satisfied, the generated parameters represent the probability distribution of the source location.

Abstract: A method for allocating parallel, independent, data tasks includes receiving data tasks, each of the data tasks having a penalty function, determining a generic ordering of the data tasks according to the penalty functions, wherein the generic ordering includes solving an aggregate objective function of the penalty functions, the method further including determining a schedule of the data tasks given the generic ordering, which packs the data tasks to be performed.

Abstract: A method for allocating parallel, independent, data tasks includes receiving data tasks, each of the data tasks having a penalty function, determining a generic ordering of the data tasks according to the penalty functions, wherein the generic ordering includes solving an aggregate objective function of the penalty functions, the method further including determining a schedule of the data tasks given the generic ordering, which packs the data tasks to be performed.

Abstract: A method, apparatus, and computer program product for scheduling stream-based applications in a distributed computer system with configurable networks are provided. The method includes choosing, at a highest temporal level, jobs that will run, an optimal template alternative for the jobs that will run, network topology, and candidate processing nodes for processing elements of the optimal template alternative for each running job to maximize importance of work performed by the system. The method further includes making, at a medium temporal level, fractional allocations and re-allocations of the candidate processing elements to the processing nodes in the system to react to changing importance of the work. The method also includes revising, at a lowest temporal level, the fractional allocations and re-allocations on a continual basis to react to burstiness of the work, and to differences between projected and real progress of the work.

Abstract: A method for allocating parallel, independent, data tasks includes receiving data tasks, each of the data tasks having a penalty function, determining a generic ordering of the data tasks according to the penalty functions, wherein the generic ordering includes solving an aggregate objective function of the penalty functions, the method further including determining a schedule of the data tasks given the generic ordering, which packs the data tasks to be performed.

Abstract: A method, an apparatus and an article of manufacture for processing a random-walk based vertex-proximity query on a graph. The method includes computing at least one vertex cluster and corresponding meta-information from a graph, dynamically updating the clustering and corresponding meta-information upon modification of the graph, and identifying a vertex cluster relevant to at least one query vertex and aggregating corresponding meta-information of the cluster to process the query.

Abstract: Techniques for scheduling one or more MapReduce jobs in a presence of one or more priority classes are provided. The techniques include obtaining a preferred ordering for one or more MapReduce jobs, wherein the preferred ordering comprises one or more priority classes, prioritizing the one or more priority classes subject to one or more dynamic minimum slot guarantees for each priority class, and iteratively employing a MapReduce scheduler, once per priority class, in priority class order, to optimize performance of the one or more MapReduce jobs.

Abstract: Techniques for scheduling multiple flows in a multi-platform cluster environment are provided. The techniques include partitioning a cluster into one or more platform containers associated with one or more platforms in the cluster, scheduling one or more flows in each of the one or more platform containers, wherein the one or more flows are created as one or more flow containers, scheduling one or more individual jobs into the one or more flow containers to create a moldable schedule of one or more jobs, flows and platforms, and automatically converting the moldable schedule into a malleable schedule.

Abstract: Techniques for scheduling multiple flows in a multi-platform cluster environment are provided. The techniques include partitioning a cluster into one or more platform containers associated with one or more platforms in the cluster, scheduling one or more flows in each of the one or more platform containers, wherein the one or more flows are created as one or more flow containers, scheduling one or more individual jobs into the one or more flow containers to create a moldable schedule of one or more jobs, flows and platforms, and automatically converting the moldable schedule into a malleable schedule.

Abstract: Techniques for determining feasibility of a set of one or more operator partitioning constraints are provided. The techniques include receiving one or more sets of operator partitioning constraints, wherein each set of one or more constraints define one or more desired conditions for grouping together of operators into partitions and placing partitions on hosts, wherein each operator is embodied as software that performs a particular function, processing each set of one or more operator partitioning constraints to determine feasibility of each set of one or more operator partitioning constraints, creating and outputting one or more candidate partitions and one or more host placements for each set of feasible partitioning constraints, and creating and outputting a certificate of infeasibility for each set of infeasible partitioning constraints, wherein the certificate of infeasibility outlines one or more reasons for infeasibility.

Abstract: Techniques for partitioning an operator flow graph are provided. The techniques include receiving source code for a stream processing application, wherein the source code comprises an operator flow graph, wherein the operator flow graph comprises a plurality of operators, receiving profiling data associated with the plurality of operators and one or more processing requirements of the operators, defining a candidate partition as a coalescing of one or more of the operators into one or more sets of processing elements (PEs), using the profiling data to create one or more candidate partitions of the processing elements, using the one or more candidate partitions to choose a desired partitioning of the operator flow graph, and compiling the source code into an executable code based on the desired partitioning.

Abstract: A method for scheduling a data processing job includes receiving the data processing job formed of a plurality of computing units, combining the plurality of computing units into a plurality of sets of tasks, each set including tasks of about equal estimated size, and different sets having different sized tasks, and assigning the tasks to a plurality of processors using a dynamic longest processing time (DLPT) scheme.

Abstract: Techniques for partitioning an operator flow graph are provided. The techniques include receiving source code for a stream processing application, wherein the source code comprises an operator flow graph, wherein the operator flow graph comprises a plurality of operators, receiving profiling data associated with the plurality of operators and one or more processing requirements of the operators, defining a candidate partition as a coalescing of one or more of the operators into one or more sets of processing elements (PEs), using the profiling data to create one or more candidate partitions of the processing elements, using the one or more candidate partitions to choose a desired partitioning of the operator flow graph, and compiling the source code into an executable code based on the desired partitioning.

Abstract: An apparatus and method for making fractional assignments of processing elements to processing nodes for stream-based applications in a distributed computer system includes determining an amount of processing power to give to each processing element. Based on a list of acceptable processing nodes, a determination of fractions of which processing nodes will work on each processing element is made. To update allocations of the amount of processing power and the fractions, the process is repeated.

Abstract: Techniques for partitioning an operator flow graph are provided. The techniques include receiving source code for a stream processing application, wherein the source code comprises an operator flow graph, wherein the operator flow graph comprises a plurality of operators, receiving profiling data associated with the plurality of operators and one or more processing requirements of the operators, defining a candidate partition as a coalescing of one or more of the operators into one or more sets of processing elements (PEs), using the profiling data to create one or more candidate partitions of the processing elements, using the one or more candidate partitions to choose a desired partitioning of the operator flow graph, and compiling the source code into an executable code based on the desired partitioning.

Abstract: Techniques for scheduling a plurality of jobs sharing input are provided. The techniques include partitioning one or more input datasets into multiple subcomponents, analyzing a plurality of jobs to determine which of the plurality of jobs require scanning of one or more common subcomponents of the one or more input datasets, and scheduling a plurality of jobs that require scanning of one or more common subcomponents of the one or more input datasets, facilitating a single scanning of the one or more common subcomponents to be used as input by each of the plurality of jobs.

Abstract: A method, an apparatus and an article of manufacture for processing a random-walk based vertex-proximity query on a graph. The method includes computing at least one vertex cluster and corresponding meta-information from a graph, dynamically updating the clustering and corresponding meta-information upon modification of the graph, and identifying a vertex cluster relevant to at least one query vertex and aggregating corresponding meta-information of the cluster to process the query.

Abstract: A system and method for choosing non-continual jobs to run in a stream-based distributed computer system includes determining a total amount of resources to be consumed by non-continual jobs. A priority threshold is determined above which jobs will be accepted, below which jobs will be rejected. Overall penalties are minimized relative to the priority threshold based on estimated completion times of the jobs. System constraints are applied to ensure that jobs meet set criteria such that a plurality of non-continual jobs are scheduled which consider the system constraints and minimize overall penalties using available resources.

Abstract: A method of choosing jobs to run in a stream based distributed computer system includes determining jobs to be run in a distributed stream-oriented system by deciding a priority threshold above which jobs will be accepted, below which jobs will be rejected. Overall importance is maximized relative to the priority threshold based on importance values assigned to all jobs. System constraints are applied to ensure jobs meet set criteria.