Intel Corporation

Alexander V Supalov

Vladimir Dmitrievich Truschin

Alexey V Ryzhykh

Sergey I Sapronov

Igor Ermolaev

Andrey Semin

Michael Vyacheslavovich Chuvelev

Alexey V Bayduraev

Dmitry Mishura

William R Magro

David K Poulsen

Rob F Van Der Wijngaart

Artyom A Syrov

Dmitry V Ezhov

Andrey B Derbunovich

Dmitriy Vitalievich Dontsov

Artem V Yalozo

Stanley J Whitlock

Denis V Nagorny

Generally, this disclosure provides systems, devices, methods and computer readable media for launching MPI-based processes and applications on multiple platform types. A method may include determining a type of an OS associated with a processing node platform on which the MPI-based process is to be launched. The method may also include launching a proxy module on the processing node platform, the proxy module configured to launch the MPI-based process. The launching of the proxy module may include operations that are based on the determination of the OS type of the processing node platform. These operations may include the execution of a secure shell for remote process launching and/or requesting a remote process launching from a system service module executing on the processing node platform.

Startup of message-passing-interface (MPI) based applications in a heterogeneous environment

Illustrative embodiments of methods, machine-readable media, and parallel computing systems that automatically tune a communication protocol are disclosed. In some embodiments, a method may include selecting a collective operation to be implemented by a Message Passing Interface (MPI) of a parallel computing system, wherein the MPI is configured to implement the collective operation using at least one of a plurality of algorithms; modeling each of the plurality of algorithms as a corresponding plurality of point-to-point operations; determining a transfer rate of each of the plurality of algorithms using its corresponding plurality of point-to-point operations; selecting one or more algorithms from among the plurality of algorithms, the one or more selected algorithms having higher transfer rates than one or more non-selected algorithms; and testing performance of the one or more selected algorithms with a plurality of configuration parameters to determine at least one optimized algorithm for implementing the collective operation.

Message passing interface tuning using collective operation modeling

In a data processing system that is executing a parent execution entity of an application, the parent execution entity has a first affinity setting. The data processing system enables the parent execution entity to create a worker execution entity that has a second affinity setting without changing the affinity setting of the parent execution entity. Workload for the application may then be performed in parallel by the parent execution entity and the worker execution entity. In one embodiment, to create the worker execution entity with the second affinity setting, the system first creates a delegate execution entity that has the first affinity setting. The system then changes the affinity setting of the delegate execution entity to the second affinity setting. The delegate execution entity then creates the worker execution entity with the second affinity setting. Another embodiment involves a super-delegate execution entity. Other embodiments are described and claimed.

Methods and apparatus for controlling affinity for execution entities

Methods, systems and apparatuses for processor selection in multi-processor systems are disclosed. An example method includes, for each of a plurality of processors, retrieving a list of interrupt instances for a plurality of interrupt types; calculating an interrupt instance count value for each of the plurality of interrupt types; multiplying a corresponding weighting factor by the interrupt instance count value for each one of the plurality of interrupt types to generate a plurality of weighted interrupt values; calculating an overall weighted vector value based on the sum of the plurality of weighted interrupt values; and designating one of the plurality of processors as a selected processor based on the lowest overall weighted vector value.

Methods, systems and apparatuses for processor selection in multi-processor systems

In one embodiment, the present invention includes a system that can optimize message passing by, at least in part, automatically determining a minimum number of fabrics and virtual channels to be activated to handle pending connection requests and data transfer requests, and preventing processing of new connection requests and data transfer requests outside of a predetermined communication pattern. Other embodiments are described and claimed.

Extended dynamic optimization of connection establishment and message progress processing in a multi-fabric message passing interface implementation

Disclosed herein are systems, methods and storage medium associated with native cloud computing. In embodiments, a system may include a number of clusters of computing nodes, and a data communication network configured to couple the clusters of computing nodes. The system may further include a control node configured to segment or cause segmentation of the data communication network to isolate a cluster of the computing nodes from other clusters of the computing nodes, t for allocation for native execution of a computation task. The system may further include a control network coupled to the data communication network and the control node. Other embodiments may be disclosed and claimed.

Native cloud computing via network segmentation

In one embodiment, the present invention includes a method to obtain topology information regarding a system including at least one multicore processor, provide the topology information to a plurality of parallel processes, generate a topological map based on the topology information, access the topological map to determine a topological relationship between a sender process and a receiver process, and select a given memory copy routine to pass a message from the sender process to the receiver process based at least in part on the topological relationship. Other embodiments are described and claimed.

Optimizing memory copy routine selection for message passing in a multicore architecture

A method, system, and computer program product containing instructions for automatically converting an MPI source code program into an MPI thread-based program. In response to inputs in the form of an MPI source code program and a command, a converter declares a global variable of the MPI source code program as a thread private variable to create a first private variable for a first thread and a second private variable for a second thread. A library is identified to support converting processes to threads during execution of the MPI thread-based program, and the identified library is used to build an executable version of the MPI thread-based program. The identified library may include code to identify instantiation of a new process when the MPI thread-based program is executing, and in response, to cause a corresponding thread for the MPI thread-based program to be instantiated.

Automatic conversion of MPI source code programs into MPI thread-based programs

A method, system, and computer program product containing instructions for establishing and maintaining multiple connections over different communication fabrics between two processes. The slowest, most reliable connection may be established first and then complemented by progressively faster connections between the same pair of processes. Each of these multiple connections is maintained throughout the duration of the communication session between the processes. These multiple connections may include connections made via network interfaces and, when available, direct connections such as a shared memory connection or a point-to-point processor interconnection. This connection strategy provides one or more failback communication paths that can be used with no startup costs in the event of failure of one of the other communication paths. These failback communication paths can be used to exchange failover protocol information needed to resend messages that were undelivered due to failure of one of the communication connections.

Transparent failover support through pragmatically truncated progress engine and reversed complementary connection establishment in multifabric MPI implementation

A technique includes using a virtual channel between a first process and a second process to communicate messages between the processes. Each message contains protocol data and user data. All of the protocol data is communicated over a first channel associated with the virtual channel, and the user data is selectively communicated over at least one other channel associated with the virtual channel.

Virtual heterogeneous channel for message passing

In one embodiment, the present invention includes a method for receiving an application linked against a first application binary interface (ABI), providing an ABI wrapper associated with the application, and binding the application to a native message passing interface (MPI) library using the ABI wrapper and the profiling message passing interface (PMPI). Other embodiments are described and claimed.

Using message passing interface (MPI) profiling interface for emulating different MPI implementations

A method for memory registration caching comprising enumerating a first process for a first process, in response to a memory registration cache being activated; finding an import table for the first module, wherein one or more pointers, in the import table, refer to memory management routines in a first library; changing the pointers so that the pointers refer to memory management routines in a second library; overloading routines that refer to the memory management routines in the first library so that the routines refer to the memory management routines in the second library; intercepting memory allocation requests, wherein the size of the request is forwarded to the memory registration cache; and de-registering freed memory from the memory registration cache.

Growing community of inventors

Erftstadt, Germany

Alexander V Supalov