Nvidia Corporation

Amazon Technologies, Inc.

A9.com, Inc.

Scott Michael Le Grand

Rejith George Joseph

Saurabh Dileep Baji

Tin-Yu Lee

Bandish N Chheda

Peter Sirota

Nachiketh Rao Potlapally

John Merrill Phillips

Nicholas Patrick Wilt

Deepak Singh

Kiuk Chung

Edward Kandrot

Methods, systems, and computer-readable media for implementing a fault-tolerant parallel computation framework are disclosed. Execution of an application comprises execution of a plurality of processes in parallel. Process states for the processes are stored during the execution of the application. The processes use a message passing interface for exchanging messages with one other. The messages are exchanged and the process states are stored at a plurality of checkpoints during execution of the application. A final successful checkpoint is determined after the execution of the application is terminated. The final successful checkpoint represents the most recent checkpoint at which the processes exchanged messages successfully. Execution of the application is resumed from the final successful checkpoint using the process states stored at the final successful checkpoint.

Fault-tolerant parallel computation

Memory management is provided for processors, such as GPUs used to process data using a trained machine learning model. Requests received to a CPU can be stored to a request queue until the queue is full, or until a timeout value has been reached for periods of lower activity. The requests can then be batched and sent to a GPU as a single message on a single thread. Memory can be pre-allocated, and the trained model loaded into GPU memory once for processing of the relevant batches. The individual requests can be processed by the GPU and the results analyzed to determine at least a subset of results to return to the CPU, which can be provided back as results of the processing.

Batch processing of requests for trained machine learning model

At the request of a customer, a distributed computing service provider may create multiple clusters under a single customer account, and may isolate them from each other. For example, various isolation mechanisms (or combinations of isolation mechanisms) may be applied when creating the clusters to isolate a given cluster of compute nodes from network traffic from compute nodes of other clusters (e.g., by creating the clusters in different VPCs); to restrict access to data, metadata, or resources that are within the given cluster of compute nodes or that are associated with the given cluster of compute nodes by compute nodes of other clusters in the distributed computing system (e.g., using an instance metadata tag and/or a storage system prefix); and/or restricting access to application programming interfaces of the distributed computing service by the given cluster of compute nodes (e.g., using an identity and access manager).

Isolating compute clusters created for a customer

The present disclosure is directed to parallelization of artificial neural network processing by conditionally synchronizing, among multiple computer processors, either the input or output of individual operations, and by conditionally using either rows or columns of certain matrices used in the operations. The conditional processing may depend upon the relative sizes of the input and output of the specific operations to be performed. For example, if a current layer matrix of values is larger than a next layer matrix of values to be computed, then rows of a weight matrix may be used by the computer processors to compute the next layer matrix. If the current layer matrix is smaller than the next layer matrix, then columns of the weight matrix may be used by the computer processors to compute the next layer matrix.

Conditional parallel processing in fully-connected neural networks

A client may submit a job to a service provider that processes a large data set and that employs a message passing interface (MPI) to coordinate the collective execution of the job on multiple compute nodes. The framework may create a MapReduce cluster (e.g., within a VPC) and may generate a single key pair for the cluster, which may be downloaded by nodes in the cluster and used to establish secure node-to-node communication channels for MPI messaging. A single node may be assigned as a mapper process and may launch the MPI job, which may fork its commands to other nodes in the cluster (e.g., nodes identified in a hostfile associated with the MPI job), according to the MPI interface. A rankfile may be used to synchronize the MPI job and another MPI process used to download portions of the data set to respective nodes in the cluster.

Executing parallel jobs with message passing on compute clusters

A method for providing fault tolerance in a distributed file system of a service provider may include launching at least one data storage node on at least a first virtual machine instance (VMI) running on one or more servers of the service provider and storing file data. At least one data management node may be launched on at least a second VMI running on the one or more servers of the service provider. The at least second VMI may be associated with a dedicated IP address and the at least one data management node may store metadata information associated with the file data in a network storage attached to the at least second VMI. Upon detecting a failure of the at least second VMI, the at least one data management node may be re-launched on at least a third VMI running on the one or more servers.

Fault tolerance in a distributed file system

A service provider system may implement ECC-like features when executing computations on GPUs that do not include sufficient error detection and recovery for computations that are sensitive to bit errors. During execution of critical computations on behalf of customers, the system may automatically instrument program instructions received from the customers to cause each computation to be executed using multiple sets of hardware resources (e.g., different host machines, processor cores, or internal hardware resources). The service may provide APIs with which customers may instrument their code for execution using redundant resource instances, or specify parameters for applying the ECC-like features. The service or customer may instrument code to perform (or cause the system to perform) checkpointing operations at particular points in the code, and to compare intermediate results produced by different hardware resources. If the intermediate results do not match, the computation may be restarted from a checkpointed state.

Automated error detection and recovery for GPU computations in a service environment

One embodiment of the present invention sets forth a technique for efficiently performing broad phase collision detection using parallel spatial subdivision. The technique involves organizing candidate objects according to a hashed representation of each object centroid, constructing a cell identification (ID) array, sorting the cell ID array, creating a collision cell list, and traversing the collision cell list. The result is a candidate list of object groups that may collide, based on an initial assessment of spatial proximity. Whether a given pair of objects actually collides is determined by a precise narrow phase collision analysis.

System and method for reducing the complexity of performing broad-phase collision detection on GPUS

System and method for reducing the complexity of performing broad-phase collision detection on GPUs

One embodiment of the present invention sets forth a technique for synchronizing the execution of multiple cooperative thread arrays (CTAs) implementing a parallel algorithm that is mapped onto a graphics processing unit. An array of semaphores provides synchronization status to each CTA, while one designated thread within each CTA provides updated status for the CTA. The designated thread within each participating CTA reports completion of a given computational phase by updating a current semaphore within the array of semaphores. The designated thread then polls the status of the current semaphore until all participating CTAs have reported completion of the current computational phase. After each CTA has completed the current computational phase, all participating CTAs may proceed to the next computational phase.

Method for synchronizing independent cooperative thread arrays running on a graphics processing unit

One embodiment of the present invention sets forth a technique for efficiently performing N-body computations using parallel computation systems. The technique involves a first processing step whereby a force matrix is partitioned into tiles, which are assigned to a one or more thread groups for processing. An off-diagonal tile may be aligned to include no diagonal cells, while an on-diagonal tile includes diagonal cells. One approach for computing either type of tile involves assigning each row from a tile to a thread within a thread group. Each thread operates on an offset pattern to avoid access conflicts to a shared memory. A net force for each atom within an N-body system is then computed by efficiently adding constituent forces stored within the force matrix using reduction operations on the force matrix.

Growing community of inventors

Soquel, CA, United States of America

Scott Michael Le Grand