Parallel log structured file system collective buffering to achieve a compact representation of scientific and/or dimensional data

• US 9,122,700 B2
• Filed: 12/20/2012
• Issued: 09/01/2015
• Est. Priority Date: 12/20/2011
• Status: Active Grant

• Current Assignee: TRIAD National Security, LLC
• Original Assignee: Los Alamos National Security LLC (Government of the United States of America)
• Inventors: Grider, Gary A., Poole, Stephen W.
• Primary Examiner(s): NGUYEN, KIM T
• Application Number:
  US13/722,946
• Publication Number:
  US 20130159364A1
• Time in Patent Office: 985 Days
• Field of Search: 707/696, 707/826
• US Class Current: 1/1
• CPC Class Codes: G06F 16/1858 Parallel file systems, i.e. file systems supporting multiple processors
• CPC Class Codes: G06F 16/1865 Transactional file systems

Abstract

Collective buffering and data pattern solutions are provided for storage, retrieval, and/or analysis of data in a collective parallel processing environment. For example, a method can be provided for data storage in a collective parallel processing environment. The method comprises receiving data to be written for a plurality of collective processes within a collective parallel processing environment, extracting a data pattern for the data to be written for the plurality of collective processes, generating a representation describing the data pattern, and saving the data and the representation.

Claims

1. A method, implemented at least in part by one or more computing devices, for data storage in a collective parallel processing environment, the method comprising:

receiving data to be written for a plurality of collective processes within a collective parallel processing environment;

before the data is written, extracting a data pattern for the data to be written for the plurality of collective processes;

generating a representation describing the data pattern; and

saving the data and the representation of the data pattern;

wherein the representation is a single index entry, and wherein the representation describing the data pattern comprises information describing;

a number of the plurality of collective processes;

an offset in files being written when saving the data;

a total amount of data being written; and

an amount of data being written per collective process.

2. The method of claim 1, wherein the plurality of collective processes are organized into a plurality of compute nodes, and wherein the data to be written is received from each compute node.

3. The method of claim 1, wherein the data is structured data, wherein each collective process of the plurality of collective processes writes its respective data.

4. The method of claim 3, wherein the data and the representation are saved using Parallel Log-structured File System (PLFS).

5. The method of claim 1, wherein the data is unstructured data, wherein each collective process of the plurality of collective processes writes its respective data.

6. The method of claim 5, further comprising:

before extracting the data pattern, shuffling data between neighboring collective processes of the plurality of collective processes such that the data pattern is more easily extracted.

7. The method of claim 1, further comprising:

applying collective buffering techniques to create structured data across the plurality of collective processes.

8. The method of claim 1, further comprising:

determining that data to be written for the plurality of collective processes is unstructured data;

based on the determining, shuffling data between neighboring collective processes to create structured data, wherein the data pattern is extracted based on the structured data.

9. The method of claim 1, further comprising:

determining that data to be written for the plurality of collective processes is unstructured data;

based on the determining, shuffling data between neighboring compute nodes to create structured data, wherein the data pattern is extracted based on the structured data, and wherein the data is stored using PLFS in a number of physical files corresponding to the number of compute nodes.

10. The method of claim 1, further comprising:

determining that data to be written for the plurality of collective processes is unstructured data;

based on the determining, shuffling data between neighboring collective processes to create structured data, wherein the structured data for each collective process has the same number of bytes or the same number of complex structures, except for a remainder which is written by a last collective process.

11. The method of claim 1, wherein the plurality of collective processes are organized into a plurality of compute nodes, and wherein the data to be written is moved between neighboring compute nodes such that a same size of data to be written is received from each compute node, except for a last compute node from which a remainder of less than the same size of data may be received.

12. The method of claim 1, wherein the representation comprises information describing a structure of the data across the plurality of collective processes, wherein the structure of the data comprises an indication of a data type of the data.

13. A non-transitory computer-readable storage medium storing computer-executable instructions for causing a computing device to perform a method for data storage in a collective parallel processing environment, the method comprising:

receiving data to be written for a plurality of collective processes within a collective parallel processing environment;

before the data is written, extracting a data pattern for the data to be written for the plurality of collective processes;

generating a representation describing the data pattern; and

saving the data and the representation of the data pattern;

wherein the representation is a single index entry, and wherein the representation describing the data pattern comprises information describing;

a number of the plurality of collective processes;

an offset in files being written when saving the data;

a total amount of data being written; and

an amount of data being written per collective process.

14. A high-performance parallel processing computing environment comprising:

one or more computing devices comprising processing units and memory;

the one or more computing devices configured to perform operations comprising;

receiving data to be written for a plurality of collective processes within the parallel processing computing environment;

before the data is written, extracting a data pattern for the data to be written for the plurality of collective processes;

generating a representation describing the data pattern; and

saving the data and the representation of the data pattern, wherein the data and the representation are saved using Parallel Log-structured File System (PLFS);

wherein the representation is a single index entry, and wherein the representation describing the data pattern comprises information describing;

a number of the plurality of collective processes;

an offset in files being written when saving the data;

a total amount of data being written; and

an amount of data being written per collective process.

15. The high-performance parallel processing computing environment of claim 14, wherein the plurality of collective processes are organized into a plurality of compute nodes, and wherein the data to be written is received from each compute node.

16. The high-performance parallel processing computing environment of claim 14, wherein the data is structured data, wherein each collective process of the plurality of collective processes writes its respective data.

17. The high-performance parallel processing computing environment of claim 14, wherein the data is unstructured data, wherein each collective process of the plurality of collective processes writes its respective data.

18. The high-performance parallel processing computing environment of claim 17, the operations further comprising:

before extracting the data pattern, shuffling data between neighboring collective processes of the plurality of collective processes such that the data pattern is more easily extracted.

19. The high-performance parallel processing computing environment of claim 14, the operations further comprising:

determining that data to be written for the plurality of collective processes is unstructured data;

based on the determining, shuffling data between neighboring collective processes to create structured data, wherein the data pattern is extracted based on the structured data.