International Business Machines Corporation

Richard R Treumann

Bin Jia

Robert Sidney Blackmore

Chulho Kim

Hanhong Xue

Rajeev Sivaram

Su-Hsuan Huang

Ellick C Law

William G Tuel, Jr

Alison B White

Xin Chen

Ron Sivan

Guy Bashkansky

Sergei Shmulyian

Intra-node data transfer in collective communications is facilitated. A memory object of one task of a collective communication is concurrently attached to the address spaces of a plurality of other tasks of the communication. Those tasks that attach the memory object can access the memory object as if it was their own. Data can be directly written into or read from an application data structure of the memory object by the attaching tasks without copying the data to/from shared memory.

Facilitating intra-node data transfer in collective communications, and methods therefor

A method is provided in which checkpointing operations are carried out in data processing systems running multiple processes which employ shared memory in a manner which preserves data coherence and integrity but which places no timing restrictions or constraints which require coordination of checkpointing operations. Data structures within local process memory and within shared memory provide the checkpoint operation with application level information concerning shared memory resources specific to at least two processes being checkpointed. Methods are provided for establishing, restoring and releasing shared memory regions that are accessed by multiple cooperating processes.

Checkpoint/resume/restart safe methods in a data processing system to establish, to restore and to release shared memory regions

A method, system, and computer program product to transfer data between two application data structures by copying a data gather scatter program (DGSP) from an exporting process address space where a first data structure is located, to a location in shared memory visible to an importing process address space; assembling a parameter set identifying the data structure; starting a data gather scatter-redistribution machine (DGS-RM) in an importing process space where a second application data structure is located; passing the first parameter set, the DGSP copy, and a second parameter set identifying a second application data structure and a second DGSP to the DGS-RM; and creating master and worker stack machines. The master stack machine identifies a contiguous chunk of the first data structure. The worker stack machine identifies contiguous chunks of the second data structure representing the same number of bytes as the contiguous chunk of the first data structure and transferring to (from) one or more identified chunks of the second data structure from (to) the single chunk of the first application data structure.

Data gather scatter&#x2014;redistribution machine

A facility is provided for communicating among processes in a symmetric multi-processing (SMP) cluster environment wherein at least some SMP nodes of the SMP cluster include multiple processes. The facility includes transferring intra-nodal at an SMP node messages of a collective communication among processes employing a shared memory of the SMP node; and responsive to the intra-nodal transferring, concurrently transferring inter-nodal multiple messages of the collective communication from n SMP node(s) to m other SMP node(s), wherein at least one of n or m is greater than one. The concurrently transferring is performed by multiple processes of at least one of the n SMP node(s) or the m other SMP node(s). More particularly, the facility includes concurrently transferring inter-nodal the multiple messages from one of: one SMP node to multiple other SMP nodes, multiple SMP nodes to one other SMP node, or multiple SMP nodes to multiple other SMP nodes.

Method, system and program product for communicating among processes in a symmetric multi-processing cluster environment

A method, an apparatus and a recording medium are provided for communicating message payload data, especially noncontiguous message data, from a first node of a network to a second node of the network in response to a request to transmit a message. Such method includes dividing the length of a data payload to be transmitted into a plurality of submessage payload lengths, i.e., into at least a first submessage payload length and a second submessage payload length. Then, a first ordered submessage is transmitted from the first node for delivery to the second node, the first ordered submessage having the first submessage payload length. A first state of an environment is then determined in the first node as if the step of transmitting the first ordered submessage were already completed. Without having to complete the step of transmitting the first ordered submessage, a second ordered submessage is then transmitted from the first node for delivery to the second node, the second submessage having the second submessage payload length, the second submessage being transmitted in a way that takes into account the first state of the environment in the first node.

Method and apparatus for striping message payload data over a network

A method for implementing Message Passing Interface (MPI-2) one-sided communication by using Low-level Applications Programming Interface (LAPI) active messaging capabilities, including providing at least three data transfer types, one of which is used to send a message with a message header greater than one packet where Data Gather and Scatter Programs (DGSP) are placed as part of the message header; allowing a multi-packet header by using a LAPI data transfer type; sending the DGSP and data as one message; reading the DSGP with a header handler; registering the DSGP with the LAPI to allow the LAPI to scatter the data to one or more memory locations; defining two sets of counters, one counter set for keeping track of a state of a prospective communication partner, and another counter set for recording activities of local and Remote Memory Access (RMA) operations; comparing local and remote counts of completed RMA operations to complete synchronization mechanisms; and creating a mpci_wait_loop function.

Method for implementing MPI-2 one sided communication

Shared locks are employed for controlling a thread which extends across more than one protocol layer in a data processing system. The use of a counter is used as part of a data structure which makes it possible to implement shared locks across multiple layers. The use of shared locks avoids the processing overhead usually associated with lock acquisition and release. The thread which is controlled may be initiated in either an upper layer protocol or in a lower layer.

Sharing lock mechanism between protocol layers

Facilitating intra-node data transfer in collective communications

A method and system for transferring noncontiguous messages group including assembling a set of data into a series of transmission packets, packaging a description of the layout of the transmission packets into description packets and then places each description packet into a local buffer while maintaining a count of the number of description packets, transfers each description packet into a transmit buffer for transmission to at least one receiving node, identifies the data packets, and forwards each data packet to the transmit buffer for transmission to the at least one receiving node. The receiving node receives the transmission packets, identifies each packet as a description packet or data packet, places the description packets in a local buffer for storage until the description is complete, places each description packet into a user data buffer, stores data packets in a local queue until the description is complete, then transfers the data packets to the user buffer.

Method and system for efficiently transferring a self-defined non-contiguous message in a one-sided communication model

An embodiment of the present invention is directed to a method for compiling, storing, and interpreting, as often as needed, a representation of any MPI datatype, including the steps of compiling a tree representation of an MPI datatype into a compact, linear data gather scatter program (DGSP) wherein the DGSP is of a form general enough to encode an arbitrarily complex datatype, registering the compact linear DGSP with a communications subsystem for later interpretation by the subsystem for at least one of sends, receives, packs and unpacks, creating a registered DGSP, and interpreting the registered DGSP. In one embodiment of the present invention, the form of the DGSP uses a single generalized representation. In another embodiment the single generalized representation covers any of the arbitrarily complex datatype patterns that can arise in this context. In yet another embodiment, the single generalized representation provides that any datatype that can be constructed using an application programming interface (API) in MPI can be converted into the form.

Data gather/scatter machine

A deterministic, non-deadlocking technique to achieving distributed consensus in a multithreaded multiprocessing computing environment is provided. A communicator is established across multiple processes in the multithreaded computer environment notwithstanding that multiple groups of threads may be simultaneously trying to establish communicators. The technique includes communicating across the multiple processes to establish a candidate identifier for the communicator for a group of participating threads of the multiple processes; and communicating across the multiple processes to check at each participating thread of the multiple processes whether the candidate identifier can be claimed at its process, and if so, claiming the candidate identifier as the new identifier thereby establishing the communicator. As one example, the technique can be implemented via a subroutine call within a message passing interface (MPI) library.

Growing community of inventors

Highland, NY, United States of America

Richard R Treumann