2. Agenda
• The problem – exploding Data
• What is heterogeneous computing
• How does it work
• Some success stories
• DB as a choke point
• impact on the enterprise
• Nagios
• Our experience – some data
• Melt Iron – who we are.
3. Problem – exploding Data…
2.5 Exabytes/day generated
• Google handles > 24 PB
• 7 billion shares
• Giga (109) …Tera …Peta …Exa
…Zetta …Yottabytes
Implications for Compute
• CAPEX follows curve
• OPEX follows curve
Implication
• Compute capability will not match
growing data.
4. Types of Data
Transactional Data / MIS, Business Apps.
• Decision making for each transaction.
• Key metric: Transactions per sec.
• Acceleration - Faster processors & in-memory compute.
Parallel data / Big-Data / Analytics.
• Pattern analysis
• Math/Statistical analysis
• Key metric: data size stored & processed
• Acceleration - Parallel Processor.
Interdependent data / Weather Forecasting, HPC
• Scenario planning, Scientific computing
• Finite element analysis
5. Heterogenous computing (CPU+GPU)
•CPU is sequential computing
•GPU is parallel computing.
•General computation on CPU
•Parallel data on GPU
•CPU has 4-8-32 cores, GP-GPUs have 1K-5K cores
• Analogy: a truck vs. a Freight Train
•Truck carries small load & relatively flexible
•Train carries huge load and is relatively inflexible
6. GPU Based Computing
• GPU – Graphics processing unit. Parallel
cores, used for graphics, video, streaming
media.
• GP-GPU: General purpose GPUs - used in
high performance computing (HPC) for
very large data sets.
• Offload data intensive processing to GPU,
rest to CPU.
• Power efficient data centers, Govt. Labs,
Universities, Large enterprises use GP-
GPU.
• Performance improvements of 50x and
more.
7. GP-GPU stories*
BNP Paribas- 10x lower power, 16x lesser space
J.P. Morgan :Risk Computation
40x performance Improvement
80x lower data centre costs
Bloomberg-:Fixed Income 16hrs-2hrs
Bond Valuation 8x faster
38x lower energy costs
AON Benfield :Insurance-Risk Management
From days to minutes-can respond in intra day now
Citadel: Hedge Fund
70x faster pricing
* from NVIDIA
9. Nagios: Architecture
DB slows as records added.
5 GB limit
New Apps: Analytics, ML
Answers:
• cluster, partition/shard db
• modify query/apps
• expensive and done post-deployment…
11. Our experience: GP-GPU usage
Pattern match on:
• Zeon quad-core server – 8 GB
• nVidia Quadro 2000: 192 cores, 1 GB RAM
• As data size increases, CPU slows exponentially
• GPU curve is almost flat.
12. Melt Iron
• We are about parallel-computing
• Focused on the Enterprise.
• Huge, huge opportunity everywhere in Enterprise Compute.
• Change the course of the river Amazon – from sequential to
heterogeneous compute
• Open Source
• Will setup meetup on heterogeneous computing
• Welcome open source contributors:
• Java/C++
• C/Asm
• CUDA/OpenCL
contact me: rinka@meltiron.com
13. Melt Iron: DB appliance
Web Server
JDBC/ODBC
Java/C# Enterprise App…
Web Server
JDBC/ODBC
Java/C# Enterprise App…
Web Server
JDBC/ODBC
Java/C# Enterprise App…
Database
Melt Iron
DB Appliance (HA)
Web Server
JDBC/ODBC
Real-time Analytics App…
Accelerate DB by
more than 100x.