New Business Applications Powered by In-Memory Technology @MIT Forum for Supply Chain Innovation, 2011

New Business Applications
Powered by In-Memory
Technologies from Academia
MIT Forum for Supply Chain Innovation
HPI February 2011
Paul Hofmann, PhD
VP, Group of the Chief Scientist
SAP Chief
Scientist
Group

Examples of Projects with In-Memory
Technology in Collaboration with Academia
© SAP AG 2011. All rights reserved. / Page 2
Taking In-Memory Computing Seriously
Coherent Shared Memory – BigIron in Palo Alto
Keep programming model simple AND solve very complex problems
Princeton University
 Optimal pricing for energy management, online pricing, truck scheduling, …
Infinite DRAM - RAMCloud
Stanford University and HPI
 Extremely low latency and very high bandwidth
 Facebook like problems with high read AND write rate
 Advanced analytics, what-if scenarios, demand planning, ...
Hybrid In-Memory Store
MIT CSAIL and HPI
 Aggregate column store – the best of both worlds
Multithreading Real Time Event Platform
MIT Auto-ID Lab and HPI
 500k events/s and millions of threads in-memory or distributed
 Automatic meter reading, online billing, mobile billing, Smart Grid

Taking In-Memory Computing Seriously
Chief Scientist Group
Princeton University – Operations Research and Financial Engineering
Warren Powell, et al

Taking In-Memory Computing Seriously!
Basic Assumptions
 Disk is tape - active data must be in DRAM
 Data locality is king  avoid cache misses and stalled CPU
Problems and Opportunities for In-Memory Computing
 Addressable DRAM per box is limited – different than hard disks.
We need to scale memory independently from physical boxes
 Scaling Architecture
– Arbitrary scaling of the amount of data stored in DRAM
– Arbitrary & independent scaling of number of active users & associated computing load
 Inter-Process Communication is slow and hard to program (latencies are in the area of
0.5-1ms )
We can do better

Taking In-Memory Computing Seriously!
How can we do better?
 Coherent Shared Memory or ccNUMA
All CPUs can access all memory and all I/O channels in about 1 μs
We can scale independently with DRAM and CPUs
You need more computing power – add another board …
You need more DRAM – add another board …
 Merge Application Server & DB Server – reference memory directly from app

BigIron - A System We Architected For Hana with
Leading-Edge, Cluster Server Components
System Specifications Architecture, Assembly,
System architecture: SAP Technology
Infrastructure
Research Practice
Assembly and Test: Colfax
International
Hosting: Bay Area Internet
Solutions, Santa
Clara, CA
• Large shared
coherent
memory (5TB)
across
servers via
Scale MP
• 160 cores(320
HT)
Big Iron 2
Extreme Performance, Scalability,
and much simpler system model
Research Server Cluster
 5 x 4U Servers
(4 Intel XEON x7560 2.26Ghz)
 160 cores (32 Cores/Server)
 5TB memory (64 x 16MB
DDR3/Server)
 30TB SSD (solid state disk)
storage
 5 Networks
 VPN of ScaleMP (40-
160GbIB)
 VPN of Server Cluster
(10GbE)
 VPN of Storage Array
(10GbE)
 VPN of SAP Internal Network
(10MbE metered)
 Firewalled GW to Internet
(1GbE Expandable)
 1 NAS (72TB Expandable to 180)
 1 x 48U Rack
 System Software
 SLES11 Linux OS Licenses
 ScaleMP vSMP Licenses
 System cost: $618K with tax and
support
& Hosting

Coherent Shared Memory – Improved
Productivity at Low Price
Traditional server clusters – Distributed memory
Programming of Queries,
Distribution of Data
Physical
Server #1
Server Cluster with Coherent Shared Memory
Traditional
SAP developer
Coherent Shared Memory via
software approach
 Developers can treat system as one “big” server and
let the operating system and lower level software
handle the problem
 Initial design is timeless – hardware scaling handled
below app design layer
 Developers do not need additional skills for in-memory
computing
Before
Physical
Server #2
SAP developer
for in-memory
computing
Physical
Server n …
 Developers need to distribute queries and data
across physical servers; access to data requires
mastering complex communications protocols
 Design trapped at a single “scale” – platform growth
forces redesign every couple years
 Specialized programming skills held only by SAP’s
top developers
After
Physical
Server #1
Physical
Server #2
Physical
Server n …

Solve Very Compute Intensive Problems
Like Stochastic Optimization
We need to juggle intermittent energy from wind or solar and volatile electricity
prices to meet time-varying loads – Princeton has the necessary algorithms
Wind speed
Electricity prices
We can reduce compute time from days to minutes!
Load

Modeling uncertainty in power scheduling
12
10
8
6
4
2
The effect of modeling uncertainty in wind
0
Uncertain forecast Perfect forecast Constant wind
2% wind
40% wind

Modeling Uncertainty In Power Scheduling
Designing energy portfolios….
… is like building a stone wall. You can do a perfect job with a perfect forecast. The
challenge is dealing with uncertainty.

In the beginning….

Infinite DRAM with RAM Cloud
Stanford University and HPI
John Ousterhout, Mendel Rosenblum, Christian Tinnefeld, et al

Impact of Latency for Internet Applications
Large-scale Apps Struggle with High Latency
Web Application
Application Server Storage Server
0.5-10ms latency

RAMCloud – Stanford University
Create distributed storage system that keeps data entirely in DRAM
Combining the main memories of thousands of servers
Use high-end network (10 Gigabit Ethernet / Infiniband)
Replicate data synchronously into the main memories of other nodes
Asynchronous writes to disk only for backup/archival purpose
This results in:
 Gracefully scaling storage solution
 Very low access latency AND high bandwidth
 Latency of memory access via network 1-5 μs
 Eventually consistent data storage – consistency is sub second

RAMCloud – Stanford University
RamCoud ideal for apps with millions of concurrent users that need more complex
data structures than key value store e.g. Facebook, Quora, Yelp, etc.
Research Questions in the Context of Enterprise Applications
What-If Analytics
 Ideally, companies want to do what-if analytics on their complete history of transactional
data, not on a subset (Think about WalMart or Unilever)
 What-if Analytics need high bursts of read AND write access
 How can the needed data be modeled and placed in a RAMCloud?
 How can the needed data operations be implemented in a scalable way?
Transactional Properties
 Enterprise applications rely on transactional properties
 RAMCloud provides extremely low latency, but only eventually consistent
 Are the reduced transaction times sufficient to avoid lock wait times or reduce aborted
transactions?

Hybrid In-Memory Store HIRYSE
MIT CSAIL and HPI
Sam Madden, Philippe Cudre-Mauroux, Jens Krueger, Martin Grund, et al

Hybrid Partitioning for Mixed Workloads
OLTP
OLAP
Row
1
Row
2
Row
3
Row
1
Row
2
Row
3
Row Store Column Store Hybrid Store
Row
4
Row
4
Doc
Num
Doc
Date
Sold-
To
Value
Sales
Org
Status
Doc
Num
Doc
Date
Sold-
To
Value
Sales
Org
Status
Row
1
Row
2
Row
3
Row
4
Doc
Num
Sold-
To
Doc
Date
Value
Sales
Org
Status

HYRISE Architecture
■ Query Processor - chooses the best
possible query plan for the hybrid data
storage structure
■ Layout Manager - given a workload
it performs an evaluation based on a
cache miss based cost model and
generates the best possible layout
■ Storage Manager - main memory
hybrid storage manager capable of
vertical partitioning in single relations

Cost Model
Used to describe the layout dependent costs based on cache misses
Combine complex access patterns from simpler ones (scan, lookup,…)
Measure
 All different access variants to the hybrid storage layer - full projection, partial
projection, selection, tuple reconstruction
 Different performance parameters - compiler settings, hardware pre-fetcher, …
Observe
 Combined behavior of all components – look at total workload
 Count cache misses using the CPU hardware performance counter
Understand
 Develop the cost model

Results

Multithreading Real Time Event Platform
MIT Auto ID Lab and HPI
John Williams, Sergio Herrero, Abel Sanchez, et al

Motivation: Rapid Growth of Events and
Messaging Platforms
Verizon and T-Mobile: 2-3 days to generate phone bill
A Comparative Study of Data Storage and Processing Architectures
iTunes: 24 hours to generate bill
Uninterrupted Growth of online billing systems (Hulu, Netflix…)
Dynamic Pricing on SmartGrid requires design of infrastructure
capable of ingesting millions of events in quasi-real time
Goal: Design a
multi-threaded
system that
produces the
electricity
consumption bill
of a city of 1M
households
8 hours  seconds

Smart Meter Reading Problem
Users
Energy
Producers
Data Generation
Data Persistence
Data Processing

Multithreaded Simulator for Smart Meter
Management
Layer
Generation Layer Consumption Layer
Storage Layer
Meter Read
Generators
Head Ends
Simulation
Manager
Query
Client
Query
Client
Query
Client
MDUS
Meter
Data
Unification
System
DRYAD
HADOOPDB
SAP HANA
Web Service
Interfaces
Web Service
Interfaces

Conclusion
Platform that handles billions of events/day AND large numbers of
threads on one machine (> 1 million), e.g. Siemens 500k events/s
RDBMS (used by today’s MDUS vendors) provides good query
performance but does not scale to millions of households (8 h)
Distributed File System (DFS): Provides the scalability, reliability & insert
performance necessary for the storage of Smart Meter Reader data
Using Map Reduce on top of a DFS (HDFS or CEPH) is good for batch
processing systems
Loading data from DFS and executing queries in-memory using HANA
provides very good performance results for real time queries  RAPTOR
Prototype for SmartGrid allowing to ingest smart meter data in real time,
do dynamic pricing (4 buckets), store in DFS & do real time analytics
A Comparative Study of Data Storage and Processing Architectures
Bill for 1 M households in seconds

Sense
• Collect current conditions at
fine grain in real time
Analyze
• Access real-time data – analyze
data, learn, build models
Actuate
• Control (soft, hard, persuasive,
personal)
ELECTRONIC NERVOUS SYSTEM
Analytical
Approach
Inductive
Approach
No Data Complete Data
No Prior
Knowledge
Perfect
Knowledge
Data
Knowledge
2
  
   
2
1
exp
2
i
i
x
y

 
 
 
X1 -1 Y1 0.02540487
X2 -0.9 Y2 0.02779527
X3 -0.8 Y3 0.03010825
X4 -0.7 Y4 0.03228947
X5 -0.6 Y5 0.03428442 .. .. .. .. . . . .
GPS SIM
Card

New Business Applications Powered by In-Memory Technology @MIT Forum for Supply Chain Innovation, 2011

More Related Content

What's hot (20)

Viewers also liked (9)

Similar to New Business Applications Powered by In-Memory Technology @MIT Forum for Supply Chain Innovation, 2011 (20)

Recently uploaded (20)

New Business Applications Powered by In-Memory Technology @MIT Forum for Supply Chain Innovation, 2011