1 / 10

High Throughput Compression of Double-Precision Floating-Point Data

High Throughput Compression of Double-Precision Floating-Point Data. Martin Burtscher and Paruj Ratanaworabhan School of Electrical and Computer Engineering Cornell University. Introduction. Scientific programs Produce and transfer lots of 64-bit FP data

marrim
Download Presentation

High Throughput Compression of Double-Precision Floating-Point Data

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. High Throughput Compression of Double-Precision Floating-Point Data Martin Burtscher and Paruj Ratanaworabhan School of Electrical and Computer Engineering Cornell University

  2. Introduction • Scientific programs • Produce and transfer lots of 64-bit FP data • Exchange 100s of MB/s, generate 1TB/day of new data • Large amounts of data • Are expensive to store and transfer • Take a long time to transfer • Data compression • Can reduce amount of data • Can speed up transfer March 2007

  3. IEEE 754 Double-Precision Values • Goal • Compress linear streams of FP data fast and well • Online operation and lossless compression • Challenges • Floating-point data are hard to compress • FP codes may generate over 90% unique values • Related work on lossless FP compression • Focuses on 32-bit single-precision values • Relies on smoothness of data or known geometry March 2007

  4. Floating-Point Data Compression • Our approach • Predict FP data with value prediction algorithms and encode the difference • Format: • Value predictors • Hardware devices to speed up processors • Predict instruction result by extrapolating previously sequences of computed results • Employ very fast and simple algorithms March 2007

  5. FPC Algorithm • Make two predictions • Select closer value • XOR with true value • Count leading zeros • Encode value • Update predictors March 2007

  6. Algorithm/Implementation Co-Design • Inner loop (about 50 and 70 C statements) • Compresses or decompresses one block of data • Accounts for over 90% of execution time • Loop body optimizations • Loop body is used to hide memory latency • No fp, int mult, or int div instructions • No branches (only conditional moves) • Single basic block (>100 machine instructions) • Average IPC > 5.4 and 5.1 on Itanium 2 March 2007

  7. Evaluation Method • System • 1.6 GHz Itanium 2, Intel C Itanium Compiler 9.1 • Red Hat Enterprise Linux AS4 • Scientific datasets • Linear streams of 64-bit FP data (18 – 277MB) • 4 observations: spitzer, temp, error, info • 4 simulations: comet, plasma, brain, control • 5 messages: bt, lu, sp, sppm, sweep3d March 2007

  8. Compression Throughput March 2007

  9. Decompression Throughput March 2007

  10. Summary and Conclusions • FPC algorithm • Highest throughput and mean compression ratio • 1.02 – 15.05 absolute compression ratio • 840 and 680 MB/s throughput on a 1.6GHz Itanium 2 (= 2 and 2.5 machine cycles per byte) • http://www.csl.cornell.edu/~burtscher/research/FPC/ • Conclusions • Value predictors are fast & accurate data models • Algorithm/implementation co-design is essential March 2007

More Related