Fat Nodes & GPGPUs - Red-shifting your infrastructure without breaking the budget

2,758 views
2,563 views

Published on

A short overview on how GPGPUs and Fat Nodes can provide an approach to deal with exponential growth demands in computing that is found not just in major companies such as Google or YouTube but in many major Irish companies as well.

This hardware/software approach can accelerate your processing infrastructure whilst reducing its costs (CapEx and OpEx) and size.

Published in: Technology, Business
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
2,758
On SlideShare
0
From Embeds
0
Number of Embeds
224
Actions
Shares
0
Downloads
0
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide
  • 1 / 10 the hardware cost 1 / 20 the electrical cost
  • Fat Nodes & GPGPUs - Red-shifting your infrastructure without breaking the budget

    1. 1. Red-shifting your infrastructure without breaking the budget – Fat Nodes & GPGPUs Irish Centre for High End Computing Dr. Eoin Brazil, Technology Transfer
    2. 2. Outline of Presentation <ul><li>Red-shift versus Blue-shift </li></ul><ul><li>Scaling – Up or Out – Why not both ? </li></ul><ul><li>The secret recipe and how to bake it </li></ul><ul><li>Capacity planning </li></ul><ul><li>Rock stars wanted </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    3. 3. Red-Shift / Blue-shift Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs Rapidly growing compute and / or data demand beyond Moore’s Law
    4. 4. Characteristics of a Red-shift Company <ul><li>Rapidly growing compute and / or data demand beyond Moore’s Law </li></ul><ul><li>Not just Google or YouTube </li></ul><ul><li>Irish examples include Paddy Power, Tullow Oil, MET Éireann, NewBay and many more! </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    5. 5. Google – A Red-shift Company <ul><li>&quot;Between 1998 and 2009 it is reported that Google's infrastructure scaled performance (queries processed per day) by 1,000 times while scaling its infrastructure by 1,000 times .&quot; </li></ul><ul><li>J. Dean, “Challenges in Building Large-Scale Information Retrieval Systems,” keynote talk, Proc. 2nd Ann. ACM Conf. Web Search and Data Mining (WSDM 09), ACM Press, 2009; http://wsdm2009.org/proceedings.php </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    6. 6. YouTube – A Red-shift Problem <ul><li>Estimates that 24 hours of video are being uploaded per minute </li></ul><ul><li>HD rate is 2 – 5 Mbps that equals about 45-75 terabytes of data per day </li></ul><ul><li>Transcoding is a typical compute dependant task </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    7. 7. Scaling – Which Way ? <ul><li>Up – Bigger server or capability per server </li></ul><ul><li>Out – More servers </li></ul><ul><li>Hybrid – Do both </li></ul><ul><li>NextIO, SuperMicro, Appro, etc. </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    8. 8. Off the Shelf <ul><li>NextIO vCORE Extreme, 3u, 16 double-wide GPU’s and 8 server connections over x16 PCIe – 8 Tflops DP / 16 Tflops SP </li></ul><ul><li>SuperMicro SuperServer 7046GT-TRF-FC407, 4u tower, 4 double-wide GPU’s </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    9. 9. Big Data - Hadoop MapReduce <ul><li>Black Scholes option pricing uses Monte Carlo methods – MITHRA, ~2008/9 </li></ul><ul><li>Cluster: 62 CPU Quad Core Cluster = 248 cores </li></ul><ul><li>Machine: 4 CPU Quad Core Cluster (16 cores) + 2x 9800 GX2 GPUs (256 cores) = 272 cores </li></ul><ul><li>~2008/9, 1/10 the size and 508x speed up </li></ul><ul><li>~2011, x2 CPU versus x4 GPU improvement </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    10. 10. Recipe for FAT Nodes <ul><li>Keep processing local </li></ul><ul><li>Use dual 12 Core CPUs with 64GB Ram on each CPU (124 GB per node) </li></ul><ul><li>Connect 10+ Fermi GPUs to each node </li></ul><ul><li>Use SSDs (1 SSD @ 200K = 50 ~ 100 HDD in parallel on single PCIe card) </li></ul><ul><li>Connect with 40Gb/s InfiniBand </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    11. 11. Scaling & Optimising <~> Capacity Planning Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs Benchmark real runs against Amdahl's Law and Neil J. Gunther’s Universal Scalability Law (USL) to give insight into future planning for system and/or infrastructure.
    12. 12. Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    13. 13. Training on CUDA <ul><li>For Academics: </li></ul><ul><ul><li>FREE, HPC, MPI, OpenMP and CUDA </li></ul></ul><ul><li>For Businesses: </li></ul><ul><ul><li>COMMERICAL RATES, all the same courses plus custom tailoring </li></ul></ul><ul><ul><li>Consultancy </li></ul></ul><ul><li>NVIDIA Certified Courses </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    14. 14. <ul><li>5 forthcoming new positions at ICHEC </li></ul><ul><li>http://www.ichec.ie/about_us/employment </li></ul><ul><li>Email us: [email_address] </li></ul>Irish Centre for High End Computing (ICHEC) - Optimisation with GPGPUs
    15. 15. Acknowledgements Supported by Science Foundation Ireland under grant 08/HEC/I1450 and by HEA’s PRTLI-C4. GPGPU - Gilles Civario, ICHEC

    ×