1 / 25

SDQuery DSI: Integrating Data Management Support with a Wide Area Data Transfer Protocol

SDQuery DSI: Integrating Data Management Support with a Wide Area Data Transfer Protocol. Yu Su*, Yi Wang*, Gagan Agrawal*, Rajkumar Kettimuthu # *The Ohio State University # The University of Chicago and Argonne National Laboratory. Motivation.

tamber
Télécharger la présentation

SDQuery DSI: Integrating Data Management Support with a Wide Area Data Transfer Protocol

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. SDQuery DSI: Integrating Data Management Support with a Wide Area Data Transfer Protocol Yu Su*, Yi Wang*, Gagan Agrawal*, Rajkumar Kettimuthu# *The Ohio State University #The University of Chicago and Argonne National Laboratory

  2. Motivation • Science becomes increasingly data driven • Strong requirements for efficient data analysis • “Big Data” Challenge: • Fast data generation speed • Slow disk I/O and network speed • Some number from road-runner EC3 simulation • 40003 particles, 36 bytes per particle => 2.3 TB • Network Bandwidth: GB level or even less • Huge difference between simulation and network • Gap will become bigger in future

  3. Wide-Area Data Transfer Protocols • Efficient data transfers over wide-area network • Globus GridFTP: • Striped, Streaming, Parallel Data Transfer • Reliable and Restartable Data Transfer • Limitation: volume? • The basic data transfer unit is file (GB or TB Level) • Strong requirements for transferring data subsets • Climate Simulation, Tomography, XPCS • An Example • Goal: Integrate core data management functionality with wide-area data transfer protocols

  4. Challenges • How should the method be designed to allow easy use and integration with existing GridFTP installation? • How can users view a remote file and conveniently specify the subsets of data that is of interest to them? • How to support efficient data retrieval with different subsetting scenarios (index-based retrieval or data block loading + in-memory filter)? • How can data retrieval be parallelized and benefits from multi-steaming?

  5. Introduction • GridFTP SDQuery DSI (Scientific Data QueryData Storage Interface) • Efficient Data Transfer over Flexible File Subset • Dynamic Loading / Unloading • HDF5 and NetCDF Data Formats • Standard SQL Embedded in Data Download Request • Multiple Query Types (Dims, Coordinates, Values) • Bitmap Indexing • Metadata View of Data File • Features: • Performance Model based Hybrid Data Reading • Parallel Streaming Data Reading and Transferring

  6. Background: GlobusGridFTP • Support Efficient Data Transfer in Grid Community • 3500+ server, 1PB+ transfer/day • DSI(Data Storage Interface): • Compatible with different file systems or platforms • An adapter between GridFTP and system • SDQuery DSI: • Dynamic loading with small overhead • Seamless integration with GridFTP data transfer features (Fault Tolerance, Security, Automatic TCP optimization)

  7. Background: Bitmap Indexing • Widely used in scientific data management • Suitable for float value by binning small ranges • Run Length Compression(WAH, BBC) • Compress bitvector based on continuous 0s or 1s

  8. System Architecture GridFTP Client GridFTP Client GridFTP Client GridFTP Server Request Parser data store request schema request data retrieve request Parse SQL query Query Analysis File Receiver Receive Data File File Receiver File Reader Indexing and find all data pos Index Operations Build Multi-level Bitmap Indexing Index Generation File DSI Read Data based on data pos Data Reader Generate Metadata View Schema Management Send File Query Metadata View File Sender SDQuery DSI Indices and schema HDF5, NetCDF Dataset

  9. Metadata View

  10. An User Case POP.nc TEMP(Query).nc Less Than 5% Data Transfer! • Translate Analysis Requirement into Query: • Find the data elements under the depth of 50 meters of the ocean and the temperature is larger than 5 centigrade.

  11. Performance Model-based Data Subset Retrieval • Data Retrieval Process: • Query Analysis and Index Operations - Fast • Know how much data to fetch after index operations: • Data Reader – Slow • Data Reading Choices: • Direct Access: Smaller Data Subset • Directly read data by points or segments from disk • Memory Filter: Bigger Data Subset • Load the data blocks into memory and filter • Which method is more efficient to choose is tricky • Execution Environment, Data Format and Dataset

  12. Performance Model • Profiling and formulate data reading • Memory Filter: • Direct Access (Points): • Direct Access(Segments): • Offline Training based on random query set • Parameters are trained and classified based on subset percent • Apply formulas for each real query • Select more efficient methods for data reading

  13. Parallel Streaming • Multi-Thread Data Retrieval and Transfer: • Data retrievals are performed in parallel • Data transfers are performed in parallel to better utilize the bandwidth • Data retrievals and data transfers are performed in a pipeline mode • Bit-1 distribution based data partition: • Partition result bitset based on thread number • Great load balance for both data retrieval and transfer • Small partition cost • One pass for both bits segmenting and partition • Use multi-thread to speedup

  14. Parallel Streams Example (2 streams) T10: reading… T10: reading… Dim-based Partition Bit1-based Partition One pass: Generate Segs and Count Chunk0 Chunk1 Chunkn …… T11: waiting… T11: sending… Subset Size: 12 Subset Size: 8 Sending Queue 1 TCP stream Load Imbalance T21: waiting… T21: sending… Sending Queue 2 TCP stream Subset Size: 9 Subset Size: 5 Chunk0 Chunk1 Chunkn …… T20: reading… T20: reading…

  15. Experiment Results • Goals: • Compare SDQuery DSI with GridFTP default File DSI • Show the effectiveness of perform-model based selection between direct access and memory filter • Speedup for using parallel streaming data transfer • Datasets: • NetCDF: Parallel Ocean Programs (POP) • HDF5: Mediterranean Ocean Data Base (MODB) • Environment: • RI Cluster: 100 nodes, 8 cores 2.53 GHz Intel(R) Xeon Processors, 12 GB memory

  16. SDQuery DSI vs. File DSI • Compare the total execution time between two DSIs in different network environments • File DSI (GridFTP default DSI): • Read the entire data file and transfer over network • Dataset: • 140 GB POP data file • TEMP.nc(time(10), depth(42), lat(2400), lon(3600)) • Three Network Environment: • LAN: 1 Gb/s bandwidth, 0.17 msec RTT • WAN: Avg. 200 Mb/s bandwidth, 24 msec RTT • WAN: Avg. 20Mb/s bandwidth, 60 msec RTT

  17. SDQuery vs. File DSI (1Gb) • Data file: 140 GB • Input of SDQuery DSI: • 2000 queries cover different data subset percentage • When the data subset percentage is <50%, SDQuery DSI is better, the speedup is 1.26 to 9.41 • Otherwise: FileDSI achieves better efficiency • SDQuery Query Processing Time: Query parsing and bitmap indexing time • SDQuery Subset and Transfer Time: Data subset fetching and transfer time • File Read and Transfer Time: Entire data file reading and transfer time

  18. SDQuery vs. File DSI (200 Mb) • Same data and same input • Network transfer time becomes the main bottleneck • SDQuery DSI: Query Process Time: 9% - 40% of Total Execution Time • Compared to File DSI, SDQuery DSI achieves better efficiency for all cases. The speedup is from 1.15 to 29.07 • SDQuery Query Processing Time: Query parsing and bitmap indexing time • SDQuery Subset and Transfer Time: Data subset fetching and transfer time • File Read and Transfer Time: Entire data file reading and transfer time

  19. SDQuery vs. File DSI (20 Mb) • In a common wide area network environment where bandwidth is really limited. • Network transfer time becomes the dominant factor • SDQuery DSI: Query Process Time: 1% - 9% of Total Execution Time • SDQuery DSI achieves better efficiency for all cases. The speedup is from 1.21 to 81.32 • SDQuery Query Processing Time: Query parsing and bitmap indexing time • SDQuery Subset and Transfer Time: Data subset fetching and transfer time • File Read and Transfer Time: Entire data file reading and transfer time

  20. Accuracy of Performance Model • Data Access (points): frequent data seeking, inefficient • Data Access (segments): average seg length: 300.36, speedup: 1.64 – 3.93 • Memory Filter: Similar for all different cases • Data Access (segments) and Memory Filter method achieve same performance when subset percentage is around 62% • Hybrid Access: right choice in most case (except 60% - 70%) • X axis: data subset percentage • Y axis: only data subset reading time • Direct Access, Memory Filter

  21. Speedup Using Parallel Streaming • Benefits: • Parallel TCP Streams • Parallel Data Retrieval • Data Retrieval and transfer overlap • Dataset: 10.5 GB MODB • Network Speed: 200Mb/s • 1 Steam allows the overlap between data retrieval and data transfer, the speedup is 1.19 – 1.52 compared with non overlapping • Maximum speedup using 4 streams: 1.57 – 1.75 • Bandwidth is fully utilized • X axis: data subset percentage • Y axis: data retrieval and transfer time • Non-overlapping: data is sent back only after all subset is loaded into memory

  22. Conclusion ‘‘Big Data’’ issue brings challenges for scientific data management SDQuery DSI: a GridFTP plug-in to support flexible data subsetting over HDF5 and NetCDF Seamless integration with GridFTP server Performance model based data retrieval method Parallel steaming data retrieval and transfer

  23. Contact Us If You’re Interested! Yu Su Email: su1@cse.ohio-state.edu

  24. Thanks

  25. An Example of Ocean Simulation More Efficient! I want to analyze TEMP within Indian Ocean! Data Subset Network POP.nc Entire Data File TEMP SALT UVEL VVEL GridFTP Server Back

More Related