Abstract: A method for executing an application over a plurality of nodes in each of which an application monitor and a runtime are executing includes executing a first portion of the application by first threads of the runtime of the first node and a second portion of the application by second threads of the runtime of the second node, and under control of the application monitors of the first and second nodes and while executing the first portions and second portions of the application, migrating workloads of one or more of the first threads from the first node to the second node for execution by the second threads.

Abstract: Thread local storage is allocated to a thread that is executed in different link domains that share a memory address space by initially allocating a thread local storage having a first base address in the shared memory address space to the thread and a thread local storage having a second base address in the shared memory address space to a second thread that is created as a watcher thread of the first thread. When it is determined that the first thread has made a transition from executing code from the first link domain to executing code from the second link domain, the thread local storage having the second base address is allocated to the first thread. Thereafter, when it is determined that the first thread has resumed executing code from the first link domain, the thread local storage having the first base address is allocated to the first thread.

Abstract: A coherence protocol applied to memory pages maintains coherence between memory spaces on a plurality of nodes so that the threads of the runtime are operable on any of the nodes. The nodes operating according to the coherence protocol track a state and an epoch number for each memory page residing therein. The states include a modified state in which only one particular node has an up-to-date copy of the memory page, an exclusive state in which only one particular node owns the memory page, a shared state in which all nodes that have the memory page in the shared state have the same copy, and a lost state in which the memory page cannot be either read or written. The epoch number is a number that is incremented each time the page enters the modified state and is used to determine whether the page contains data that is stale.

Abstract: Example methods and systems for elastic provisioning of container-based graphics processing unit (GPU) nodes are described. In one example, a computer system may monitor usage information associated with a pool of multiple container-based GPU nodes. Based on the usage information, the computer system may apply rule(s) to determine whether capacity adjustment is required. In response to determination that capacity expansion is required, the computer system may configure the pool to expand by adding (a) at least one container-based GPU node to the pool, or (b) at least one container pod to one of the multiple container-based GPU nodes. Otherwise, in response to determination that capacity shrinkage is required, the computer system may configure the pool to shrink by removing (a) at least one container-based GPU node, or (b) at least one container pod from the pool.

Abstract: A coherence protocol applied to memory pages maintains coherence between memory spaces on a plurality of nodes so that the threads of the runtime are operable on any of the nodes. The nodes operating according to the coherence protocol track a state and an epoch number for each memory page residing therein. The states include a modified state in which only one particular node has an up-to-date copy of the memory page, an exclusive state in which only one particular node owns the memory page, a shared state in which all nodes that have the memory page in the shared state have the same copy, and a lost state in which the memory page cannot be either read or written. The epoch number is a number that is incremented each time the page enters the modified state and is used to determine whether the page contains data that is stale.

Abstract: At least one application of a client executes via system software on a hardware computing system that includes at least one CPU and at least one coprocessor. A virtualization layer establishes unified memory address space between the client and the hardware computing system, which also includes memory associated with the at least one coprocessor. The virtualization layer then synchronizes memory associated with the client and memory associated the at least one coprocessor. The virtualization layer may be installed and run in a non-privileged, user space, without modification of the application or of the system software running on the hardware computing system.

Abstract: A pre-emptive scheduling of workloads enables improved sharing of resources of a cluster of hosts. The steps of this pre-emptive scheduling method include: adjusting priority of active workloads that are each running on one of the nodes and idle workloads that have been suspended; determining that a priority of a first workload, which is one of the idle workloads, exceeds a priority of a second workload, which is one of the active workloads and is executing on a first node of the cluster of nodes; and suspending the second workload and resuming the first workload to run on the first node.

Abstract: Execution of multiple execution streams is scheduled on at least one coprocessor. A software layer located logically between applications and the at least one coprocessor intercepts a first API call from an application and determines that a first execution stream is to be executed. Before scheduling the first execution stream, the software layer transmits a response to the application indicating that the at least one coprocessor is ready to execute another execution stream. The software layer intercepts a second API call from the application and determines that a second execution stream including one or more kernels is to be executed. The software layer determines that the one or more kernels does not have a dependency on the first execution stream. The software layer schedules the one or more kernels for execution prior to when the at least one coprocessor has completed execution of the first execution stream.

Abstract: A method for a dynamic linker to load and run an application that is executed over a plurality of nodes, includes relocating a primary binary of the application from an initial location to an executable location, loading library dependencies, altering a system call table used during execution of the application for the dynamic linker to catch all system calls made by the application, and executing the relocated primary binary from the executable location.

Abstract: A method for executing an application over a plurality of nodes in each of which an application monitor and a runtime are executing includes executing a first portion of the application by first threads of the runtime of the first node and a second portion of the application by second threads of the runtime of the second node, and under control of the application monitors of the first and second nodes and while executing the first portions and second portions of the application, migrating workloads of one or more of the first threads from the first node to the second node for execution by the second threads.

Abstract: A method for a dynamic linker to load and run an application that is executed over a plurality of nodes, includes relocating a primary binary of the application from an initial location to an executable location, loading library dependencies, altering a system call table used during execution of the application for the dynamic linker to catch all system calls made by the application, and executing the relocated primary binary from the executable location.

Abstract: A method for handling system calls during execution of an application over a plurality of nodes including a first node and a second node, includes receiving a system call from a thread running on the first node, determining that executing the system call involves resources present on the second node, sending the system call and arguments of the system call to the second node for the second node to execute the system call, receiving the results of the system call from the second node, and returning the results of the system call to the thread.

Abstract: A method for allocating and using file descriptors for an application executing over a plurality of nodes, each having a file system, includes receiving a system call from the application running on a first node to access a file in a file system, determining whether the file resides in a file system of a first node or the second node, and upon determining that the file resides on the second node, sending the system call and arguments of the system call to the one of the second nodes for execution on the one of the second nodes and returning a result of the system call executed on the second node to the application on the first node.

Abstract: Execution of multiple execution streams is scheduled on at least one coprocessor. A software layer located logically between applications and the at least one coprocessor intercepts a first API call from an application and determines that a first execution stream is to be executed. Before scheduling the first execution stream, the software layer transmits a response to the application indicating that the at least one coprocessor is ready to execute another execution stream. The software layer intercepts a second API call from the application and determines that a second execution stream including one or more kernels is to be executed. The software layer determines that the one or more kernels does not have a dependency on the first execution stream. The software layer schedules the one or more kernels for execution prior to when the at least one coprocessor has completed execution of the first execution stream.

Abstract: Execution of multiple execution streams is scheduled on a plurality of coprocessors. A software layer located logically between applications and the coprocessors determines dependencies within the execution streams, each said dependency being a condition in one of the execution streams that must be satisfied in order for execution of at least one other of the execution streams to proceed on corresponding ones of the coprocessors. The dependencies are then represented in a data structure and an optimized execution schedule is determined for the execution streams according to the dependencies. Simultaneous execution of a plurality of the execution streams is then dynamically reordered according to the optimized execution schedule.

Abstract: At least one application of a client executes via system software on a hardware computing system that includes at least one CPU and at least one coprocessor. A virtualization layer establishes unified memory address space between the client and the hardware computing system, which also includes memory associated with the at least one coprocessor. The virtualization layer then synchronizes memory associated with the client and memory associated the at least one coprocessor. The virtualization layer may be installed and run in a non-privileged, user space, without modification of the application or of the system software running on the hardware computing system.

Abstract: Execution of multiple execution streams is scheduled on a plurality of coprocessors. A software layer located logically between applications and the coprocessors determines dependencies within the execution streams, each said dependency being a condition in one of the execution streams that must be satisfied in order for execution of at least one other of the execution streams to proceed on corresponding ones of the coprocessors. The dependencies are then represented in a data structure and an optimized execution schedule is determined for the execution streams according to the dependencies. Simultaneous execution of a plurality of the execution streams is then dynamically reordered according to the optimized execution schedule.

Abstract: In accordance with an embodiment, a network device includes a network controller and at least one network interface coupled to the network controller that includes at least one media access control (MAC) device configured to be coupled to at least one physical layer interface (PHY). The network controller may be configured to determine a network path comprising the at least one network interface that has a lowest power consumption and minimum security attributes of available media types coupled to the at least one PHY.

Abstract: Multicast transmissions are efficient but do not allow for individual acknowledgement that the data was received by each receiver. This is not acceptable for isochronous systems that require specific levels of QoS for each device. A multimedia communications protocol is provided that uses a novel multi-destination burst transmission protocol in multimedia isochronous systems. The transmitter establishes a bi-directional burst mode for multicasting data to multiple devices and receiving Reverse Start of Frame (RSOF) delimiters from each multicast-destination receiver in response to multiple SOF delimiters, thus providing protocol-efficient multi-destination acknowledgements.

Abstract: Multicast transmissions do not allow for individual receivers to acknowledge that data was received by each receiver in the network. This is not acceptable for isochronous systems that require specific levels of QoS for each device. A multimedia communications protocol supports using multicast transmissions (one-to-many) in multimedia isochronous systems. A transmitter establishes a Multi-ACKed Multicast protocol within which a group of receiving devices can acknowledge the multicast transmission during a multi-acknowledgment period.