Microsoft Technology Licensing, LLC

Amazon Technologies, Inc.

A9.com, Inc.

Trishul Amit Madhukar Chilimbi

Sumit Gulwani

Kapil Vaswani

Santosh Rajagopalan

Narayanan Sadagopan

Yida Wang

Mu Li

George Karypis

Aditya Nori

Shuai Zheng

Vishwanathan Swaminathan

Hyokun Yun

Zhen Zhang

Justin Chiu

Sigmund Isy Cherem

Sankalp Nayak

Choon Hui Teo

Iman Keivanloo

A resource set which includes multiple servers with a respective plurality of training computing devices is identified for training a machine learning model. The resource set is subdivided into partition groups, such that each partition group can store a respective replica of state information of the model. The model is trained using the partition groups. The training comprises a multi-stage gathering of a portion of the state information at training computing devices of a particular partition group. Different types of communication channels between training computing devices are used in respective stages of the gathering, including inter-server communication channels in one stage and an intra-server communication channel during another stage. A trained version of the model is stored.

Distributed training of machine learning models

Devices and techniques are generally described for search using a micro-partitioned catalog. In some examples, a search system can determine a plurality of items in an online catalog. The search system can generate a similarity graph based on vector representations of the plurality of items and determine an item as a representative item for a cluster of items. The search system can receive a query and determine the cluster of items as a candidate cluster based on a comparison between the query and the representative item. The search system can determine a candidate item of the cluster of items based on a comparison between the query and the cluster of items.

Micro-partitioning based search

This paper describes preferential path profiling, which enables profiling a specified subset of all possible program paths with very low overhead. Preferential path profiling compactly identifies paths of interest using an array. More specifically, PPP assigns a unique and compact path index identifier to all interesting paths that can be used to index into a path array. The path array contains a second path value identifier that is used to distinguish interesting paths from other program paths This path numbering allows the implementation of preferential path profiling to use array-based counters instead of hash table-based counters for identifying paths of interest and gathering path profiles, which significantly reduces execution time and computational resource overhead during profiling.

Preferential path profiling

Locks which protect data structures used within atomic sections of concurrent programs are inferred from atomic sections and acquired in a manner to avoid deadlock. Locks may be inferred by expression correspondence using a backward inter-procedural analysis of an atomic section. Locks may be sorted according to a total order and acquired early in an atomic section to prevent deadlock. Multiple granularity of locks are determined and employed. Fine grained locks may be inferred and acquired to reduce contention. Coarse grained locks may be determined and substituted for fine grained locks when necessary for unbounded locations or to reduce the number of finer grained locks.

System using backward inter-procedural analysis for determining alternative coarser grained lock when finer grained locks exceeding threshold

Growing community of inventors

Seattle, WA, United States of America

Trishul Amit Madhukar Chilimbi