Microsoft Technology Licensing, LLC

Igor Ostrovsky

Stephen Harris Toub

John Joseph Duffy

Mike Liddell

Huseyin Serkan Yildiz

Joseph E Hoag

Massimo Mascaro

Joe Duffy

Edward George Essey

Emad A Omara

Vance Palmer Morrison

Hazim Shafi

Michael John Liddell

William T Colburn

Zachary David Johnson

Samer El Baghdady

Pooja Nagpal

Emad Ali

Samer El Baghady

Huseyin Yildiz

Dynamically allocated thread storage in a computing device is disclosed. The dynamically allocated thread storage is configured to work with a process including two or more threads. Each thread includes a statically allocated thread-local slot configured to store a table. Each table is configured to include a table slot corresponding with a dynamically allocated thread-local value. A dynamically allocated thread-local instance corresponds with the table slot.

Dynamically allocated thread-local storage

The present invention extends to methods, systems, and computer program products for indicating parallel operations with user-visible events. Event markers can be used to indicate an abstracted outer layer of execution as well as expose internal specifics of parallel processing systems, including systems that provide data parallelism. Event markers can be used to show a variety of execution characteristics including higher-level markers to indicate the beginning and end of an execution program (e.g., a query). Inside the execution program (query) individual fork/join operations can be indicated with sub-levels of markers to expose their operations. Additional decisions made by an execution engine, such as, for example, when elements initially yield, when queries overlap or nest, when the query is cancelled, when the query bails to sequential operation, when premature merging or re-partitioning are needed can also be exposed.

Indicating parallel operations with user-visible events

A dataflow of a distributed application is visualized in a locally simulated execution environment. A scheduler receives a job graph which includes a graph of computational vertices that are designed to be executed on multiple distributed computer systems. The scheduler queries a graph manager to determine which computational vertices of the job graph are ready for execution in a local execution environment. The scheduler queries a cluster manager to determine the organizational topology of the distributed computer systems to simulate the determined topology in the local execution environment. The scheduler queries a data manager to determine data storage locations for each of the computational vertices indicated as being ready for execution in the local execution environment. The scheduler also indicates an instance of each computational vertex to be spawned and executed in the local execution environment based on the organizational topology and indicated data storage locations.

Simulation environment for distributed programs

A method of parallel processing an ordered input data stream that includes a plurality of input data elements and a corresponding plurality of order keys for indicating an ordering of the input data elements, with each order key associated with one of the input data elements, includes processing the input data stream in a parallel manner with a plurality of worker units, thereby generating a plurality of sets of output data elements. The plurality of sets of output data elements is stored in a plurality of buffers, with each buffer associated with one of the worker units. An ordered output data stream is output while the input data stream is being processed by outputting selected output data elements from the buffers in an order that is based on the order keys.

Parallel processing of an ordered data stream

A data partitioning interface provides procedure headings to create data partitions for processing data elements in parallel, and for obtaining data elements to process, without specifying the organizational structure of a data partitioning. A data partitioning implementation associated with the data partitioning interface provides operations to implement the interface procedures, and may also provide dynamic partitioning to facilitate load balancing.

Configurable partitioning of parallel data for parallel processing

A method includes producing values with a producer thread, and providing a queue data structure including a first array of storage locations for storing the values. The first array has a first tail pointer and a first linking pointer. If a number of values stored in the first array is less than a capacity of the first array, an enqueue operation writes a new value at a storage location pointed to by the first tail pointer and advances the first tail pointer. If the number of values stored in the first array is equal to the capacity of the first array, a second array of storage locations is allocated in the queue. The second array has a second tail pointer. The first array is linked to the second array with the first linking pointer. An enqueue operation writes the new value at a storage location pointed to by the second tail pointer and advances the second tail pointer.

Producer-consumer data transfer using piecewise circular queue

A parallelism policy object provides a control parallelism interface whose implementation evaluates parallelism conditions that are left unspecified in the interface. User-defined and other parallelism policy procedures can make recommendations to a worker program for transitioning between sequential program execution and parallel execution. Parallelizing assistance values obtained at runtime can be used in the parallelism conditions on which the recommendations are based. A consistent parallelization policy can be employed across a range of parallel constructs, and inside recursive procedures.

Controlling parallelization of recursion using pluggable policies

Dynamic data partitioning is disclosed for use with a multiple node processing system that consumes items from a data stream of any length and independent of whether the length is undeclared. Dynamic data partitioning takes items from the data stream when a thread is idle and assigns the taken items to an idle thread, and it varies the size of data chunks taken from the stream and assigned to a thread to efficiently distribute work loads among the nodes. In one example, data chunk sizes taken from the beginning of the data stream are relatively smaller than data chunk sizes taken towards the middle or end of the data stream. Dynamic data partitioning employs a growth function where chunks have a size related to single aligned cache lines and efficiently increases the size of the data chunks to occasionally double the amount of data assigned to concurrent threads.

Dynamic partitioning of data by occasionally doubling data chunk size for data-parallel applications

A scheduler receives a job graph which includes a graph of computational vertices that are designed to be executed on multiple distributed computer systems. The scheduler queries a graph manager to determine which computational vertices of the job graph are ready for execution in a local execution environment. The scheduler queries a cluster manager to determine the organizational topology of the distributed computer systems to simulate the determined topology in the local execution environment. The scheduler queries a data manager to determine data storage locations for each of the computational vertices indicated as being ready for execution in the local execution environment. The scheduler also indicates to a vertex spawner that an instance of each computational vertex is to be spawned in the local execution environment based on the organizational topology and indicated data storage locations, and indicates to the local execution environment that the spawned vertices are to be executed.

The present invention extends to methods, systems, and computer program products for propagating unhandled exceptions in distributed execution environments, such as clusters. A job (e.g., a query) can include a series of computation steps that are executed on multiple compute nodes each processing parts of a distributed data set. Unhandled exceptions can be caught while computations are running on data partitions of different compute nodes. Unhandled exception objects can be stored in a serialized format in a compute node's local storage (or an alternate central location) along with auxiliary details such as the data partition being processed at the time. Stored serialized exception objects for a job can be harvested and aggregated in a single container object. The single container object can be passed back to the client.

Propagating unobserved exceptions in distributed execution environments

The present invention extends to methods, systems, and computer program products for automatically optimizing memory accesses by kernel functions executing on parallel accelerator processors. A function is accessed. The function is configured to operate over a multi-dimensional matrix of memory cells through invocation as a plurality of threads on a parallel accelerator processor. A layout of the memory cells of the multi-dimensional matrix and a mapping of memory cells to global memory at the parallel accelerator processor are identified. The function is analyzed to identify how each of the threads access the global memory to operate on corresponding memory cells when invoked from the kernel function. Based on the analysis, the function altered to utilize a more efficient memory access scheme when performing accesses to the global memory. The more efficient memory access scheme increases coalesced memory access by the threads when invoked over the multi-dimensional matrix.

Automatic optimization for programming of many-core architectures

A method of resizing a concurrently accessed hash table is disclosed. The method includes acquiring the locks in the hash table. The hash table, in a first state, is dynamically reconfigured in size into a second state. Additionally, the amount of locks is dynamically adjusted based on comparing the size of the hash table in the second state to the size of the hash table in the second state.

Concurrently accessed hash table

Embodiments are directed to implementing custom operators in a query for a parallel query engine and to generating a partitioned representation of a sequence of query operators in a parallel query engine. A computer system receives a portion of partitioned input data at a parallel query engine, where the parallel query engine is configured to process data queries in parallel, and where the queries include a sequence of built-in operators. The computer system incorporates a custom operator into the sequence of built-in operators for a query and accesses the sequence of operators to determine how the partitioned input data is to be processed. The custom operator is accessed in the same manner as the built-in operators. The computer system also processes the sequence of operators including both the built-in operators and at least one custom operator according to the determination indicating how the data is to be processed.

Growing community of inventors

Bellevue, WA, United States of America

Igor Ostrovsky