Microsoft Technology Licensing, LLC

Vaibhav Sharma

Kapil Agarwal

Dharma Kiritkumar Shukla

Ramachandran Ramjee

Affan Arshad Dar

Muthian Sivathanu

Pankaj Sharma

Rimma V Nehme

Nipun Kwatra

Bhalakumaaran Erode Ranganathan

Srinidhi Viswanatha

Vishal Gandhi

Roopesh Kumar Manda

The disclosure herein describes platform-level checkpointing for deep learning (DL) jobs. The checkpointing is performed through capturing two kinds of state data: (i) GPU state (device state), and (ii) CPU state (host state). The GPU state includes GPU data (e.g., model parameters, optimizer state, etc.) that is located in the GPU and GPU context (e.g., the default stream in GPU, various handles created by the libraries such as DNN, Blas, etc.). Only a fraction of the GPU memory is copied because the checkpointing is done in a domain-aware manner. The 'active' memory contains useful data like model parameters. To be able to capture the useful data, memory management is controlled to identify which parts of the memory are active. Also, to restore the destination GPU to the same context/state, a mechanism is used to capture such state-changing events on an original GPU and replayed on a destination GPU.

Transparent pre-emption and migration for planet-scale computer

The disclosure herein describes processing deletion requests using sequencing numbers with change feed updates. When a deletion occurs on the source data store, a deletion notification is created in a change feed on the source server. The deletion notification includes a set of deletion record IDs identifying a set of records to be deleted, a tombstone sequence number (TSN) identifying a sequence of the deletion notification within a set of deletion notifications and/or a deletion sequence number (DSN). The DSN is incremented by one each time a new deletion notification is created. A deletion notification can represent deletion of a single record or a set of records. Each deletion notification is assigned a time-to-live (TTL) value. The deletion notification is deleted at expiration of the TTL. The TSN and the DSN entries are used to determine whether any deletion updates have been missed to prevent silent failures.

Processing delete requests based on change feed of updates

A system includes a plurality of servers, a control plane to determine a first partition of a plurality of devices and to determine a subset of the plurality of servers to assign as candidate servers for the first partition, and a common data store comprising a first stream and a second stream. The control plane is to store, in the first stream, a first message indicating the first partition, the candidate servers, the second stream, and a first message tag, the candidate servers elect a primary server of the first partition from the candidate servers using the first stream, and the elected primary server inserts read and write updates associated with the plurality of devices of the first partition into the second stream.

IoT partition management and load balancing

A system includes determination of a respective performance level associated with each of a plurality of endpoints assigned to a first routing pipeline, determination of a slow one of the plurality of endpoints based on the respective performance levels, and, in response to the determination of the slow one of the plurality of endpoints, instantiation of a second routing pipeline and assignment of the slow one of the plurality of endpoints to the second routing, wherein the first routing pipeline is to receive messages and to route a first plurality of the messages to the plurality of endpoints other than the slow one of the plurality of endpoints, and wherein the second routing pipeline is to receive the messages and to route a second plurality of the messages to the slow one of the plurality of endpoints.

Adaptive routing pipelines for variable endpoint performance

Multiple geo-replicated and independent IoT (Internet of Things) hubs configured as servers with storage capacities are utilized and span across various regions in the world, to each of which IoT devices can connect and transmit telemetry data. The IoT devices, configured with sensors to generate telemetry data (e.g., temperature, pressure, etc.), can fail over to another geo-replicated IoT hub (the 'target geo-replicated IoT hub') when one IoT hub experiences technical failures or the IoT device changes location to a region which is covered by the target IoT hub. Upon receiving a connection request from the IoT device, the target geo-replicated IoT hub submits a proposal to available geo-replicated IoT hubs for ownership over the IoT device, that is, receive data from and transmit messages to the IoT device. A quorum is sought so that all IoT hubs are in agreement as to a current owner of the IoT device.

Growing community of inventors

Seattle, WA, United States of America

Vaibhav Sharma