DATA PROCESSING PERFORMANCE ENHANCEMENT IN A DISTRIBUTED FILE SYSTEM

• US 20130254246A1
• Filed: 03/21/2012
• Published: 09/26/2013
• Est. Priority Date: 03/21/2012
• Status: Active Grant

• Current Assignee: Cloudera Incorporated
• Original Assignee: Cloudera Incorporated
• Inventors: Lipcon, Todd
• Granted Patent: US 9,405,692 B2
• Time in Patent Office: Days
• US Class Current: 707/823
• CPC Class Codes: G06F 12/0804 with main memory updating G06F12/0806 takes precedence
• CPC Class Codes: G06F 12/0866 for peripheral storage systems, e.g. disk cache
• CPC Class Codes: G06F 12/0871 Allocation or management of cache space
• CPC Class Codes: G06F 13/20 for access to input/output bus
• CPC Class Codes: G06F 16/182 Distributed file systems
• CPC Class Codes: G06F 16/183 Provision of network file services by network file servers, e.g. by using NFS, CIFS network file access protocols H04L67/1097
• CPC Class Codes: G06F 2212/214 Solid state disk
• CPC Class Codes: G06F 2212/603 of operating mode, e.g. cache mode or local memory mode

Abstract

Systems and methods of data processing performance enhancement are disclosed. One embodiment includes, invoking operating system calls to optimize cache management by an I/O component; wherein, the operating system calls are invoked to perform one or more of proactive triggering of readaheads for sequential read requests of a disk; purging data out of buffer cache after writing to the disk or performing sequential reads from the desk; and/or eliminating a delay between when a write is performed and when written data from the write is flushed to the disk from the buffer cache.

Claims

1. A method for enhancing performance for data processing in a distributed file system, the method, comprising:

invoking operating system calls to optimize cache management by an I/O component;

wherein, the operating system calls are invoked to perform one or more of;

proactive triggering of readaheads for sequential read requests of a disk;

purging data out of buffer cache after writing to the disk or performing sequential reads from the desk;

eliminating a delay between when a write is performed and when written data from the write is flushed to the disk from the buffer cache.

2. The method of claim 1, wherein, the cache management is optimized for reads and writes in Hbase or MapReduce.

3. The method of claim 1, wherein, the cache management is optimized for large sequential reads and writes.

4. The method of claim 2, wherein, the large sequential reads and writes include reads and writes of data chunks between 1-5 MB or 5-10 MB, or >

10 MB.

5. The method of claim 2, wherein, the large sequential reads and writes include reads and writes of data chunks greater than 500 kB.

6. The method of claim 1, wherein, the proactive triggering causes the readaheads to be deterministically performed in lieu of heuristically performed.

7. The method of claim 1, wherein, the readahead is executed by specifying a byte range of data to read from the disk in advance of the current request.

8. The method of claim 1, wherein, data size of the data that is purged from the buffer cache is configurable or dynamically adjustable.

9. The method of claim 1, wherein, data size of the data that is purged from the buffer cache is between 4-8 MB or 8-16 MB.

10. The method of claim 1, wherein, the delay is implemented by default to support re-writes of the written data.

11. The method of claim 3, wherein, the delay is eliminated in the distributed file system where re-writes are not supported.

12. The method of claim 1, wherein, the operating system calls are native to the operating system.

13. The method of claim 1, further comprising, enhancing performance of the distributed file system by decreasing checksum overhead in speeding up the distributed file system read path.

14. The method of claim 13, wherein, the checksum is performed in hardware by a processor supporting CRC32.

15. The method of claim 13, wherein, the checksum implementation is modified from thezlib polynomial to the iSCSI polynomial.

16. The method of claim 1, wherein, the operating system is Linux or Unix-based operating systems.

17. The method of claim 1, wherein, the operating system conforms to POSIX.1-2001.

18. The method of claim 18, wherein, the distributed file system is the Hadoop distributed file system.

19. A system for enhancing performance for data processing in a distributed file system, the system, comprising:

means for, invoking operating system calls to optimize cache management by an I/O component of the operating system for reads and writes in MapReduce;

means for, decreasing checksum overhead in speeding up the distributed file system read path to enhance performance of the distributed file system;

means for, optimizing the distributed file system (DFS) for random read performance.

20. The system of claim 19, wherein, the optimizing the DFS for random read performance includes, optimizing read performance for reading data that is less than 50 kB, less than 100 kB, or less than 1 MB.

21. The system of claim 19, wherein, the random read performance is optimized by holding an established connection used for one read operation for a subsequent read operation with a client.

22. The system of claim 19, wherein, the operating system calls are invoked to perform one or more of;

proactive triggering of readaheads for sequential read requests of a disk;

purging data out of buffer cache after writing to the disk or performing sequential reads from the desk;

eliminating a delay between when a write is performed and when written data from the write is flushed to the disk from the buffer cache.

23. A system for distributed computing, the system, comprising:

a set of machines forming a distributed file system cluster, a given machine in the set of machines having;

a processor;

a disk;

memory having stored there on instructions which when executed by the processor, causes;

readaheads for sequential read requests of the disk to be proactively triggered;

data to be purged out of buffer cache after writing to the disk or performing sequential reads from the desk.

24. The system of claim 23, further comprising,

a client coupled to the given machine,

wherein, random read performance by the client is optimized by holding an established connection with the given machine used for one read operation for a subsequent read operation.

25. The system of claim 24, wherein, the client and the given machine have same or similarly configured timeouts.

26. The system of claim 24, wherein, the established connection is held for 0.5-1 seconds, 1-2, seconds, or 2-5 seconds for optimization of the random read performance.

27. The system of claim 23, wherein, execution of the instructions further causes data written to be immediately flushed to the disk without a delay in the buffer cache intended for re-writes of the written data.

28. The system of claim 23, wherein, the distributed file system is the Hadoop distributed file system (HDFS).

29. The system of claim 23, wherein, the cache management is optimized for reads and writes in Hbase or MapReduce.

30. The system of claim 23, wherein, checksum is performed in hardware by the processor to decrease overhead by speeding up the distributed file system read or write paths.

31. The system of claim 30, wherein, the checksum uses CRC32C algorithm.