Abstract: Techniques herein minimally communicate between computers to repartition a graph. In embodiments, each computer receives a partition of edges and vertices of the graph. For each of its edges or vertices, each computer stores an intermediate representation into an edge table (ET) or vertex table. Different edges of a vertex may be loaded by different computers, which may cause a conflict. Each computer announces that a vertex resides on the computer to a respective tracking computer. Each tracking computer makes assignments of vertices to computers and publicizes those assignments. Each computer that loaded conflicted vertices transfers those vertices to computers of the respective assignments. Each computer stores a materialized representation of a partition based on: the ET and vertex table of the computer, and the vertices and edges that were transferred to the computer. Edges stored in the materialized representation are stored differently than edges stored in the ET.

Abstract: Techniques herein minimally communicate between computers to repartition a graph. In embodiments, each computer receives a partition of edges and vertices of the graph. For each of its edges or vertices, each computer stores an intermediate representation into an edge table (ET) or vertex table. Different edges of a vertex may be loaded by different computers, which may cause a conflict. Each computer announces that a vertex resides on the computer to a respective tracking computer. Each tracking computer makes assignments of vertices to computers and publicizes those assignments. Each computer that loaded conflicted vertices transfers those vertices to computers of the respective assignments. Each computer stores a materialized representation of a partition based on: the ET and vertex table of the computer, and the vertices and edges that were transferred to the computer. Edges stored in the materialized representation are stored differently than edges stored in the ET.

Abstract: Techniques are provided for dynamically self-balancing communication and computation. In an embodiment, each partition of application data is stored on a respective computer of a cluster. The application is divided into distributed jobs, each of which corresponds to a partition. Each distributed job is hosted on the computer that hosts the corresponding data partition. Each computer divides its distributed job into computation tasks. Each computer has a pool of threads that execute the computation tasks. During execution, one computer receives a data access request from another computer. The data access request is executed by a thread of the pool. Threads of the pool are bimodal and may be repurposed between communication and computation, depending on workload. Each computer individually detects completion of its computation tasks. Each computer informs a central computer that its distributed job has finished. The central computer detects when all distributed jobs of the application have terminated.

Abstract: Techniques herein minimally communicate between computers to repartition a graph. In embodiments, each computer receives a partition of edges and vertices of the graph. For each of its edges or vertices, each computer stores an intermediate representation into an edge table (ET) or vertex table. Different edges of a vertex may be loaded by different computers, which may cause a conflict. Each computer announces that a vertex resides on the computer to a respective tracking computer. Each tracking computer makes assignments of vertices to computers and publicizes those assignments. Each computer that loaded conflicted vertices transfers those vertices to computers of the respective assignments. Each computer stores a materialized representation of a partition based on: the ET and vertex table of the computer, and the vertices and edges that were transferred to the computer. Edges stored in the materialized representation are stored differently than edges stored in the ET.

Abstract: Techniques herein minimally communicate between computers to repartition a graph. In embodiments, each computer receives a partition of edges and vertices of the graph. For each of its edges or vertices, each computer stores an intermediate representation into an edge table (ET) or vertex table. Different edges of a vertex may be loaded by different computers, which may cause a conflict. Each computer announces that a vertex resides on the computer to a respective tracking computer. Each tracking computer makes assignments of vertices to computers and publicizes those assignments. Each computer that loaded conflicted vertices transfers those vertices to computers of the respective assignments. Each computer stores a materialized representation of a partition based on: the ET and vertex table of the computer, and the vertices and edges that were transferred to the computer. Edges stored in the materialized representation are stored differently than edges stored in the ET.

Abstract: Techniques herein provide job control and synchronization of distributed graph-processing jobs. In an embodiment, a computer system maintains an input queue of graph processing jobs. In response to de-queuing a graph processing job, a master thread partitions the graph processing job into distributed jobs. Each distributed job has a sequence of processing phases. The master thread sends each distributed job to a distributed processor. Each distributed job executes a first processing phase of its sequence of processing phases. To the master thread, the distributed job announces completion of its first processing phase. The master thread detects that all distributed jobs have announced finishing their first processing phase. The master thread broadcasts a notification to the distributed jobs that indicates that all distributed jobs have finished their first processing phase. Receiving that notification causes the distributed jobs to execute their second processing phase.

Abstract: Techniques minimize communication while loading a graph. In a distributed embodiment, each computer loads some edges of the graph. Each edge connects a source vertex (SV) to a destination vertex. For each SV of the edges, the computer hashes the SV to detect a tracking computer (TrC) that tracks on which computer does the SV reside. Each computer informs the TrC that the SV originates an edge that resides on that computer. For each SV, the TrC detects that the SV originates edges that reside on multiple providing computers (PCs). The TrC selects a target computer (TaC) from the multiple PCs to host the SV. The TrC instructs each PC, excluding the TaC, to transfer the SV and related edges that are connected to the SV to the TaC. A vertex's internal identifier indicates which computer hosts the vertex. The TrC maintains a mapping between external and internal identifiers.

Abstract: Techniques herein provide job control and synchronization of distributed graph-processing jobs. In an embodiment, a computer system maintains an input queue of graph processing jobs. In response to de-queuing a graph processing job, a master thread partitions the graph processing job into distributed jobs. Each distributed job has a sequence of processing phases. The master thread sends each distributed job to a distributed processor. Each distributed job executes a first processing phase of its sequence of processing phases. To the master thread, the distributed job announces completion of its first processing phase. The master thread detects that all distributed jobs have announced finishing their first processing phase. The master thread broadcasts a notification to the distributed jobs that indicates that all distributed jobs have finished their first processing phase. Receiving that notification causes the distributed jobs to execute their second processing phase.

Abstract: Techniques herein provide job control and synchronization of distributed graph-processing jobs. In an embodiment, a computer system maintains an input queue of graph processing jobs. In response to de-queuing a graph processing job, a master thread partitions the graph processing job into distributed jobs. Each distributed job has a sequence of processing phases. The master thread sends each distributed job to a distributed processor. Each distributed job executes a first processing phase of its sequence of processing phases. To the master thread, the distributed job announces completion of its first processing phase. The master thread detects that all distributed jobs have announced finishing their first processing phase. The master thread broadcasts a notification to the distributed jobs that indicates that all distributed jobs have finished their first processing phase. Receiving that notification causes the distributed jobs to execute their second processing phase.

Abstract: Techniques are provided for dynamically self-balancing communication and computation. In an embodiment, each partition of application data is stored on a respective computer of a cluster. The application is divided into distributed jobs, each of which corresponds to a partition. Each distributed job is hosted on the computer that hosts the corresponding data partition. Each computer divides its distributed job into computation tasks. Each computer has a pool of threads that execute the computation tasks. During execution, one computer receives a data access request from another computer. The data access request is executed by a thread of the pool. Threads of the pool are bimodal and may be repurposed between communication and computation, depending on workload. Each computer individually detects completion of its computation tasks. Each computer informs a central computer that its distributed job has finished. The central computer detects when all distributed jobs of the application have terminated.

Abstract: Techniques are provided for dynamically self-balancing communication and computation. In an embodiment, each partition of application data is stored on a respective computer of a cluster. The application is divided into distributed jobs, each of which corresponds to a partition. Each distributed job is hosted on the computer that hosts the corresponding data partition. Each computer divides its distributed job into computation tasks. Each computer has a pool of threads that execute the computation tasks. During execution, one computer receives a data access request from another computer. The data access request is executed by a thread of the pool. Threads of the pool are bimodal and may be repurposed between communication and computation, depending on workload. Each computer individually detects completion of its computation tasks. Each computer informs a central computer that its distributed job has finished. The central computer detects when all distributed jobs of the application have terminated.

Abstract: Techniques herein minimally communicate between computers to repartition a graph. In embodiments, each computer receives a partition of edges and vertices of the graph. For each of its edges or vertices, each computer stores an intermediate representation into an edge table (ET) or vertex table. Different edges of a vertex may be loaded by different computers, which may cause a conflict. Each computer announces that a vertex resides on the computer to a respective tracking computer. Each tracking computer makes assignments of vertices to computers and publicizes those assignments. Each computer that loaded conflicted vertices transfers those vertices to computers of the respective assignments. Each computer stores a materialized representation of a partition based on: the ET and vertex table of the computer, and the vertices and edges that were transferred to the computer. Edges stored in the materialized representation are stored differently than edges stored in the ET.

Abstract: Techniques minimize communication while loading a graph. In a distributed embodiment, each computer loads some edges of the graph. Each edge connects a source vertex (SV) to a destination vertex. For each SV of the edges, the computer hashes the SV to detect a tracking computer (TrC) that tracks on which computer does the SV reside. Each computer informs the TrC that the SV originates an edge that resides on that computer. For each SV, the TrC detects that the SV originates edges that reside on multiple providing computers (PCs). The TrC selects a target computer (TaC) from the multiple PCs to host the SV. The TrC instructs each PC, excluding the TaC, to transfer the SV and related edges that are connected to the SV to the TaC. A vertex's internal identifier indicates which computer hosts the vertex. The TrC maintains a mapping between external and internal identifiers.

Abstract: Techniques herein provide job control and synchronization of distributed graph-processing jobs. In an embodiment, a computer system maintains an input queue of graph processing jobs. In response to de-queuing a graph processing job, a master thread partitions the graph processing job into distributed jobs. Each distributed job has a sequence of processing phases. The master thread sends each distributed job to a distributed processor. Each distributed job executes a first processing phase of its sequence of processing phases. To the master thread, the distributed job announces completion of its first processing phase. The master thread detects that all distributed jobs have announced finishing their first processing phase. The master thread broadcasts a notification to the distributed jobs that indicates that all distributed jobs have finished their first processing phase. Receiving that notification causes the distributed jobs to execute their second processing phase.

Abstract: Techniques herein index data transferred during distributed graph processing. In an embodiment, a system of computers divides a directed graph into partitions. The system creates one partition per computer and distributes each partition to a computer. Each computer builds four edge lists that enumerate edges that connect the partition of the computer with a partition of a neighbor computer. Each of the four edge lists has edges of a direction, which may be inbound or outbound from the partition. Edge lists are sorted by identifier of the vertex that terminates or originates each edge. Each iteration of distributed graph analysis involves each computer processing its partition and exchanging edge data or vertex data with neighbor computers. Each computer uses an edge list to build a compactly described range of edges that connect to another partition. The computers exchange described ranges with their neighbors during each iteration.

Abstract: Techniques are provided for dynamically self-balancing communication and computation. In an embodiment, each partition of application data is stored on a respective computer of a cluster. The application is divided into distributed jobs, each of which corresponds to a partition. Each distributed job is hosted on the computer that hosts the corresponding data partition. Each computer divides its distributed job into computation tasks. Each computer has a pool of threads that execute the computation tasks. During execution, one computer receives a data access request from another computer. The data access request is executed by a thread of the pool. Threads of the pool are bimodal and may be repurposed between communication and computation, depending on workload. Each computer individually detects completion of its computation tasks. Each computer informs a central computer that its distributed job has finished. The central computer detects when all distributed jobs of the application have terminated.

Abstract: Techniques herein index data transferred during distributed graph processing. In an embodiment, a system of computers divides a directed graph into partitions. The system creates one partition per computer and distributes each partition to a computer. Each computer builds four edge lists that enumerate edges that connect the partition of the computer with a partition of a neighbor computer. Each of the four edge lists has edges of a direction, which may be inbound or outbound from the partition. Edge lists are sorted by identifier of the vertex that terminates or originates each edge. Each iteration of distributed graph analysis involves each computer processing its partition and exchanging edge data or vertex data with neighbor computers. Each computer uses an edge list to build a compactly described range of edges that connect to another partition. The computers exchange described ranges with their neighbors during each iteration.

Abstract: Techniques are provided for latency-hiding context management for concurrent distributed tasks. A plurality of task objects is processed, including a first task object corresponding to a first task that includes access to first data residing on a remote machine. A first access request is added to a request buffer. A first task reference identifying the first task object is added to a companion buffer. A request message including the request buffer is sent to the remote machine. A response message is received, including first response data responsive to the first access request. For each response of one or more responses of the response message, the response is read from the response message, a next task reference is read from the companion buffer, and a next task corresponding to the next task reference is continued based on the response. The first task is identified and continued.

Abstract: Techniques herein perform workload-balanced graph partitioning. Each graph partition is distributed to a respective computer. Each computer applies a workload-estimation function to its partition to calculate a numeric workload-value that indicates how much computation the partition needs. Each computer sends its numeric workload-value to a master computer. The master compares the highest and lowest numeric workload-values. If the difference exceeds a threshold, the master detects how much work should overloaded-computers offload to under-utilized computers. To each overloaded-computer, the master sends a directive with a balancing numeric workload-value that indicates how much computation to offload and an identifier of an under-utilized computer to receive the offload.

Abstract: Techniques are provided for latency-hiding context management for concurrent distributed tasks. A plurality of task objects is processed, including a first task object corresponding to a first task that includes access to first data residing on a remote machine. A first access request is added to a request buffer. A first task reference identifying the first task object is added to a companion buffer. A request message including the request buffer is sent to the remote machine. A response message is received, including first response data responsive to the first access request. For each response of one or more responses of the response message, the response is read from the response message, a next task reference is read from the companion buffer, and a next task corresponding to the next task reference is continued based on the response. The first task is identified and continued.