Microsoft Technology Licensing, LLC

Google Inc.

Wisconsin Alumni Research Foundation

Muthian Sivathanu

Dharma Kiritkumar Shukla

Rimma V Nehme

Ramachandran Ramjee

Lu Xun

Madhusudhan Talluri

Pankaj Sharma

Saurabh Goyal

Nipun Kwatra

Rajiv Mathews

Rajesh S R

Brent VerWeyst

Martin James Cochran

Puneet Garg

Atul Katiyar

Nishanth Chandran

Bhalakumaaran Erode Ranganathan

Vaibhav Sharma

Shreshth Singhal

Satyanarayana V Lokam

Srinidhi Viswanatha

Divya Gupta

Apurv Mehra

Andrea C Arpaci-Dusseau

Remzi H Arpaci-Dusseau

Sambhav Satija

Sudheesh Singanamalla

The disclosure herein describes scheduling execution of artificial intelligence (AI) workloads in a cloud infrastructure platform. A global scheduler receives AI workloads associated with resource ticket values. The scheduler distributes the AI workloads to nodes based on balancing resource ticket values. Local schedulers of the nodes schedule AI workloads on resources based on the resource ticket values of the AI workloads. Based on scheduling the AI workloads, coordinator services of the local schedulers execute the distributed AI workloads on the infrastructure resources of the nodes. The disclosure further describes scheduling AI workloads based on priority tiers. A scheduler receives AI workloads, and each AI workload is associated with a priority tier indicative of a preemption priority while being executed. The AI workloads are scheduled for execution on a distributed set of nodes based on the priority tiers and then execute based on the scheduling.

Scheduler for planet-scale computing system

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 platform-level migration for deep learning training (DLT) jobs from a checkpointed stated between a source node and a destination node. The checkpointing is performed through capturing GPU state (e.g., device state) and CPU state (e.g., 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 libraries). Restoring the DLT job on the destination node involves resumption of processing of a destination GPU at the same checkpointed state.

Artificial intelligence workload migration for planet-scale artificial intelligence infrastructure service

A system includes a set of low resource devices, each configured to receive transactions to be added to an encrypted block chain ledger from a sample of untrusted high resource devices, prepare a proposed block of the received transactions, provide the proposed block to the sample of untrusted high resource devices, receive proposed blocks from the untrusted high resource devices originating from the set of low resource devices. The low resource devices run a consensus protocol to select one proposed block to add to the encrypted block chain ledger stored on the untrusted high resource devices.

Lightweight blockchain based on split-trust

Methods, systems, and computer programs are presented for structuring a database to support multiple partitioning orders at the storage layer. One method includes an operation for identifying partitioning fields for a database that is stored distributed across computing devices, where each computing device stores an extent that holds a subset of entries from the database. For each partitioning field, the database entries are stored in extents associated with the partitioning field, the database entries in the extents for the partitioning field being organized based on the value of the partitioning field. Further, the method includes operations for receiving a database query that includes a filter based on values of a selected partitioning field, and for retrieving the data for the database query from one or more of the extents associated with the selected partitioning field. The retrieved data is returned for the database query.

Efficient multi-dimensional partitioning and sorting in large-scale distributed data processing systems

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium for access control for enterprise information. In one aspect, a method includes receiving resources of an enterprise, each resource having a respective access control list specifying access privileges to the resource for one or more members, and the resources including entities related to the enterprise and relationships; identifying entity facts of the entities from the resources; determining, for each entity fact, an entity fact access control list; storing data describing the entities, entity facts and the respective entity fact access control lists, wherein each entity fact is associated with its corresponding entity fact access control list; and providing, to each of the members of the enterprise, access privileges to the data describing the entities and the entity facts according to the respective entity fact access control lists.

Access control for enterprise knowledge

Implementations provide an indexing system with near-instant updates to an inverted index while maintaining techniques for query optimization. The system may provision empty positions in posting lists to enable in-place updating, without having to rebuild the posting list or append updates to the end of the posting list. For example, a system comprises at least one processor and memory storing an index that includes at least one posting list that maps a term to a set of the documents. The posting list includes an ordered list of documents and has a plurality of open positions within the ordered list. The memory also stores instructions that, when executed by the at least one processor, cause the system to locate an open position of the plurality of open positions for a new document and to insert the new document into the at least one posting list using the open position.

In-place updates for inverted indices

A system for efficiently responding to proximity queries may include a memory storing an index for searching a graph-based data store, the index including posting lists for one or more proximity ranges compatible with a space. A posting list can include one or more entities of a type compatible with the space, each entity having a location within the space, the location being a basic unit in a location hierarchy for the space and, for each entity, at least one node in the location hierarchy that falls within the proximity range of the posting list with reference to the location of the entity. The system may also include a memory storing instructions that cause the system to use the index to respond to a query that includes a query proximity range for the space. The space can be a geographic space or a time space.

Efficient index for low latency search of large graphs

Implementations provide an indexing system with an instant failover that uses a moving snapshot window. For example, a method may include receiving, by a processor, a query and determining that a main query processing engine is not responding. The method may further include generating a search result for the query using a secondary query processing engine that applies at least one snapshot record to a portion of a posting list, the snapshot record including the portion of the posting list as it appeared before a modification, and the modification occurring within a predetermined time before receiving the query. The portion is a fixed size smaller than the posting list. Applying the snapshot record can include overlaying the portion of the posting list with the snapshot record beginning at an offset specified by the snapshot record. The main query processing engine generates a search result without applying snapshot records.

Updating inverted indices

A search index for searching a graph-based data store can include triple entries, each triple entry having a posting list value, at least one intersection identifier associated with the posting list value, and at least one result identifier associated with the intersection identifier. The index may also include search entries having a posting list value that corresponds to a text search aid. The search index may also include pre-computed path entries, such as chain path entries and converge path entries. The index may also include bucket posting lists representing ranges of object values for a particular predicate and proximity posting lists that include one or more entities and the areas of a location hierarchy with locations within the proximity of the entity. Queries for the data graph may have at least two stages, each stage being associated with a posting list from a graph index.

Index and query serving for low latency search of large graphs

Methods for increasing the efficiency of data transfers by passing a reference to the data rather than to transfer the data itself. When a new communication begins, a memory object, e.g. a buffer, is selected to receive the data. Information if provided that may be used to ascertain the communication path without actual knowledge of the path. If a communication pattern is likely to be repeated, a memory object created on the initial access is saved. The saved memory object ('precursor') is provided to the memory system on subsequent access as a hint about the process path to be used. The memory system can select a cached buffer that has a similar set of mappings to the precursor.

Use of a precursor to select cached buffer

A disk drive or similar storage medium uses a semantic understanding of its associated file system to monitor file metadata and derive block liveness normally only known by the file system. Knowledge of block liveness can be used to improve the disk performance and to create a disk that provides for secure deletion without explicit instructions from the file system.

Computer storage device providing implicit detection of block liveness

Methods for performing zero-copy memory transfers between processes or services using shared memory without the overhead of current schemes. An IPC move semantic may be used that allows a sender to combine passing a reference and releasing it within the same IPC call. An insulate method removes all references to the original object and creates a new object pointing to the original memory if a receiver requires exclusive access. Alternatively, if a receiving process or service seeks read-only access, the sender unmaps its access to the buffer before sending to the receiver. When the insulate operation is initiated, the kernel detects an object with multiple active references but no active mappings and provides a mapping to the memory without taking a copy or copy-on-write.

Growing community of inventors

Bengaluru, India

Muthian Sivathanu