Method and apparatus for artificial intelligence acceleration

• US 10,467,527 B1
• Filed: 01/31/2018
• Issued: 11/05/2019
• Est. Priority Date: 01/31/2018
• Status: Active Grant

• Current Assignee: Pure Storage, Inc.
• Original Assignee: Pure Storage, Inc.
• Inventors: Margaglia, Fabio, Watkins, Emily, Kannan, Hari, Sandvig, Cary A.
• Primary Examiner(s): Vincent, David R
• Application Number:
  US15/885,665
• Time in Patent Office: 643 Days
• Field of Search: 706 12, 706 15, 706 45
• CPC Class Codes: G06F 16/28 Databases characterised by their database models, e.g. relational or object models
• CPC Class Codes: G06F 16/30 of unstructured textual data document management systems G06F16/93
• CPC Class Codes: G06F 3/061 Improving I/O performance
• CPC Class Codes: G06F 3/064 Management of blocks
• CPC Class Codes: G06F 3/067 Distributed or networked storage systems, e.g. storage area networks [SAN], network attached storage [NAS]
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 5/04 Inference or reasoning models

Abstract

An apparatus for artificial intelligence acceleration is provided. The apparatus includes a storage and compute system having a distributed, redundant key value store for metadata. The storage and compute system having distributed compute resources configurable to access, through a plurality of authorities, data in the solid-state memory, run inference with a deep learning model, generate vectors for the data and store the vectors in the key value store.

Claims

1. An apparatus for artificial intelligence acceleration, comprising:

a storage and compute system having a distributed, redundant key value store for metadata; and

the storage and compute system having distributed compute resources configurable to access, through a plurality of authorities, data in the solid-state memory, run inference with a deep learning model, generate vectors for the data and store the vectors in the key value store, the distributed compute resources configurable to store metadata for one or more buckets containing the data and create bucket views containing links to objects within the data that are contained in the one or more buckets and stored in the solid-state memory.

2. The apparatus of claim 1, further comprising:

the storage and compute system having at least one network port, and configurable to receive an object or file, run the inference for the object or the file with the deep learning model, based on the vectors in the key value store, and output a response based on the inference.

3. The apparatus of claim 1, wherein storage capacity of the solid-state memory is sufficient to hold all of the data for the inference with the deep learning model throughout the processing of an entire data pipeline.

4. The apparatus of claim 1, wherein the plurality of authorities are configurable for distributed file deletion in accordance with the data identifiers owned by each authority.

5. The apparatus of claim 1, wherein the links satisfy a similarity predicate, based on the vectors in the key value store.

6. The apparatus of claim 1, further comprising:

the distributed compute resources configurable to compute a vector of an object for a search, perform a hash operation on the vector, search for vectors in the key value store that are within a specified distance of the vector, based on a result of the hash operation, and retrieve one or more full vectors for objects stored in the plurality of solid-state memory storages based on a result of the search, wherein the vectors are stored in the key value store using a locality-preserving hash operation.

7. The apparatus of claim 1, further comprising:

the distributed compute resources configurable to receive tags associated with the data, and generate a list or count of data entities having tags that satisfy a predicate.

8. A method for artificial intelligence acceleration, performed by a storage and compute system, the method comprising:

accessing data in the system through a plurality of authorities;

performing inference with a deep learning model;

generating vectors for the data based on the performing;

storing the vectors in a key value store of the system;

storing metadata for one or more buckets containing examples in the unstructured data; and

generating bucket views containing links to the examples.

9. The method of claim 8, further comprising:

receiving an object or file; and

performing the inference for the object or file with the deep learning model based on vectors in the key value store.

10. The method of claim 8, further comprising:

outputting a response based on the inference.

11. The method of claim 8, wherein the system is configurable to write and receive unstructured data into solid state memory of the system.

12. The method of claim 8, wherein the key value store is a distributed key value store, across storage nodes of the system.

13. The apparatus of claim 9, wherein the links to the examples satisfy a similarity predicate, based on the search of the vectors.

14. A tangible, non-transitory, computer-readable media having instructions thereupon which, when executed by a processor, cause the processor to perform a method comprising:

accessing data in the system through a plurality of authorities;

performing inference with a deep learning model;

generating vectors for the data based on the performing;

storing the vectors in a key value store of the system;

storing metadata for one or more buckets containing examples in the unstructured data; and

generating bucket views containing links to the examples.

15. The computer readable media of claim 14, wherein the method further comprises:

receiving an object or file; and

performing the inference for the object or file with the deep learning model based on vectors in the key value store.

16. The computer readable media of claim 14, wherein the method further comprises:

outputting a response based on the inference.

17. The method of claim 14, wherein the system is configurable to write and receive unstructured data into solid state memory of the system.

18. The method of claim 14, wherein the key value store is a distributed key value store, across storage nodes of the system.

19. The computer readable media of claim 14, wherein the links to the examples satisfy a similarity predicate, based on the search of the vectors.