Presentation IS-4082 by Norbert Heusser at the AMD Developer Summit (APU13) November 11-13, 2013

1. REAL-­‐TIME
INSIGHT
IN
BIG
DATA
EVEN
FASTER
USING
HSA

2. AGENDA
WHAT
ARE
BIG
DATA
AND
PARSTREAM
TECHNICAL
ARCHITECTURE
HSA
USAGE
2
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL

3. What
are
Big
Data
and
ParStream

4. What
is
Big
Data?
COMMON
SENSE
FROM
WIKIPEDIA
“Big
data
is
a
collecRon
of
data
sets
so
large
and
complex
that
it
becomes
difficult
to
process
using
on-­‐hand
database
management
tools
or
tradiBonal
data
processing
applicaRons.
The
challenges
include
capture,
curaRon,
storage,
search,
sharing,
analysis
and
visualizaRon.”
4
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL

5. WHAT
BIG
DATA
IS
NOT
A
COMMON
MISTAKE
Big Data is NOT Storage of large datasets
5
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL

6. REAL-TIME IN BIG DATA IS A TWO-DIMENSIONAL PROBLEM
Continuous extremely fast data
load and availability
Sub-second response
times
6
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL

7. ANALYTICS
LANDSCAPE
BIG
DATA
ANALYTICS
REQUIRES
NEW
TECHNOLOGICAL
SOLUTIONS
OperaBonal
Data
Big
Data
Stream-­‐AnalyBcs
Real-­‐Time
Real-­‐Time
AnalyBcs
Complex
Event
Processing
OperaBons
AnalyBcs
Massively
parallel
(MPP)
Real-­‐Time
1
sec
10
sec
Batch-­‐AnalyBcs
OLAP
1
min
OLTP
ReporBng
Lag
Time
<
1..10
milli
sec
10..100
milli
sec
●
ParStream
In-­‐Memory
DB
Response
Rme
Gigabyte
7
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL
10
min
Map
Reduce
Batches
(NoSQL)
Terabyte
1h
Petabyte

8. PARSTREAM
IS
A
UNIQUE
PRODUCT
PARSTREAM
EMPOWERS
CUSTOMERS
TO
REALIZE
NEW
BUSINESS
OPPORTUNITIES
EVOLVING
WITH
BIG
DATA
! Analyze
and
Filter
Billions
of
Records
! Query
Data
Structures
with
1000’s
of
columns
! Get
Answers
in
Milliseconds
without
Cubes
! Get
Answers
in
Milliseconds
without
Cubes
Column
Store
! Execute
1000’s
of
Concurrent
Queries
High
Performance
Index
Scalability
In-­‐Memory
Technology
High-­‐Speed
Import
8
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL
Clustering
Clustering
Scalability
Real-­‐Rme
Queries

9. Technical
Architecture

10. ARCHITECTURE
BUILDING
BLOCKS
PARSTREAM
IS
THE
BIG
DATA
ANALYTICS
PLATFORM
BASED
ON
A
UNIQUE
HIGH
PERFORMANCE
COMPRESSED
INDEX
! Columnar
Storage
! In
Memory
Technology
! Shared
Nothing
Architecture
! Standard
Interfaces
SQL/JDBC/ODBC
C++
UDF
API
! User
Defined
FuncRons
! Unique
High
Performance
Compressed
Index
In-­‐Memory
&
Disc
Technology
MPP
10
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL
Real-­‐Time
AnalyRcs
Engine
Compressed
Index
Shared
Nothing
Fast
Columnar
Storage
ParRRoning

11. PARALLEL
ARCHITECTURE
PARSTREAM
OVERCOMES
LIMITATIONS
OF
TRADITIONAL
DW
ARCHITECTURES
Query
! STANDARD
DW
ARCHITECTURE
‒ Long
Query
RunRme
‒ Frequent
Full
Table
Scans
‒ Data
is
at
Least
1
Day
Old
Nightly
Batch
-­‐
Import
! PARSTREAM
ARCHITECTURE
‒ Each
Query
Uses
MulRple
Processor
Cores
‒ Query
execuRon
using
compressed
indices
‒ ConRnuous
Import
Assures
Timeliness
of
Data
11
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL
Query
HPCI
Parallel
Import

12. TRADITIONAL
DATABASE
QUERY
EXECUTION
STATIC
QUERY
EXECUTION
OpRmizer/
Planner
Parser
SQL-­‐Statement
Parsed-­‐Statement
12
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL
Executor
ExecuRonPlan

13. MODULAR
EXECUTION
TREE
ATOMIC
OPERATIONS
COMBINED
USING
QUEUES
ExecuBon
Tree
! Parsed
query
descripRons
are
transformed
into
execuRon
trees
sort
! OpRmizer
distributes
execuRon
operaRons
to
available
hardware
aggregate
! Data-­‐locality
and
current
load
are
used
for
allocaRon
! During
query
execuRon
opRmizer
can
re-­‐
allocate
if
beneficial
! OpRmizer
conRnuously
refines
allocaRon
based
on
past
queries
aggregaRon
aggregaRon
aggregaRon
aggregaRon
filter
filter
filter
filter
calc
calc
calc
calc
fetch
fetch
fetch
fetch
! Flow
based
execuRon
control
! Each
ExecNode
processes
blocks
of
data
! Data
transfer
between
nodes
using
queues
13
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL

14. HSA
Usage

15. ARCHITECUTRE
ALLOWS
USAGE
OF
DIFFERENT
PROCESSING
UNITS
ANY
PART
OF
THE
QUERY
MAY
BE
EXECUTED
INDIVIDUALLY
ExecuBon
Tree
! Each
atomic
operaRon
may
be
processed
using
any
available
compute
resource
sort
! Dynamic
workload
assignment
during
query
execuRon
aggregate
! Overall
workload
management
ensures
opRmal
resource
usage
aggregaRon
aggregaRon
aggregaRon
filter
filter
filter
filter
calc
calc
calc
calc
fetch
15
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL
aggregaRon
fetch
fetch
fetch

16. PROBLEMS
USING
TRADITIONAL
GPU
COMPUTE
UNITS
THE
TRANSFER
AND
COMMUNICATION
PROBLEM
! Target
scenario
Real-­‐Time
BIG
DATA
aggregaRon
filter
‒ Processing
huge
amounts
of
data
‒ Dynamically
changing
of
data
‒ InteracRve
response
Rme
! Part
of
the
data
fixed
in
GPU
memory
‒ Input
data
transferred
once
via
PCI
during
loading
‒ Transfer
of
result
via
PCI
during
execuRon
calc
fetch
aggregaRon
filter
calc
! Data
resident
in
main
memory
‒ Offload
of
computaRonal
task
to
GPU
‒ Transfer
in
and
out
via
PCI
during
execuRon
! Global
data
needs
to
be
transferred
to
GPU
too
! Global
data
needs
to
be
synchronized
! Latency
based
on
blockwise
processing
! Different
programming
models
16
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL
fetch

17. HSA
SOLVES
ALL
OUR
PROBLEMS
! No
Data
transfer
required
! Shared
page
table
support
! Coherent
memory
regions
!
User-­‐level
command
queueing
! Hardware
scheduling
! Bold
allows
uniform
programming
model
17
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL

18. DISCLAIMER
&
ATTRIBUTION
The
informaRon
presented
in
this
document
is
for
informaRonal
purposes
only
and
may
contain
technical
inaccuracies,
omissions
and
typographical
errors.
The
informaRon
contained
herein
is
subject
to
change
and
may
be
rendered
inaccurate
for
many
reasons,
including
but
not
limited
to
product
and
roadmap
changes,
component
and
motherboard
version
changes,
new
model
and/or
product
releases,
product
differences
between
differing
manufacturers,
soqware
changes,
BIOS
flashes,
firmware
upgrades,
or
the
like.
AMD
assumes
no
obligaRon
to
update
or
otherwise
correct
or
revise
this
informaRon.
However,
AMD
reserves
the
right
to
revise
this
informaRon
and
to
make
changes
from
Rme
to
Rme
to
the
content
hereof
without
obligaRon
of
AMD
to
noRfy
any
person
of
such
revisions
or
changes.
AMD
MAKES
NO
REPRESENTATIONS
OR
WARRANTIES
WITH
RESPECT
TO
THE
CONTENTS
HEREOF
AND
ASSUMES
NO
RESPONSIBILITY
FOR
ANY
INACCURACIES,
ERRORS
OR
OMISSIONS
THAT
MAY
APPEAR
IN
THIS
INFORMATION.
AMD
SPECIFICALLY
DISCLAIMS
ANY
IMPLIED
WARRANTIES
OF
MERCHANTABILITY
OR
FITNESS
FOR
ANY
PARTICULAR
PURPOSE.
IN
NO
EVENT
WILL
AMD
BE
LIABLE
TO
ANY
PERSON
FOR
ANY
DIRECT,
INDIRECT,
SPECIAL
OR
OTHER
CONSEQUENTIAL
DAMAGES
ARISING
FROM
THE
USE
OF
ANY
INFORMATION
CONTAINED
HEREIN,
EVEN
IF
AMD
IS
EXPRESSLY
ADVISED
OF
THE
POSSIBILITY
OF
SUCH
DAMAGES.
ATTRIBUTION
©
2013
Advanced
Micro
Devices,
Inc.
All
rights
reserved.
AMD,
the
AMD
Arrow
logo
and
combinaRons
thereof
are
trademarks
of
Advanced
Micro
Devices,
Inc.
in
the
United
States
and/or
other
jurisdicRons.
SPEC
is
a
registered
trademark
of
the
Standard
Performance
EvaluaRon
CorporaRon
(SPEC).
Other
names
are
for
informaRonal
purposes
only
and
may
be
trademarks
of
their
respecRve
owners.
18
|
REAL-­‐TIME
INSIGHT
IN
BIG
DATA|
November
19,
2013
|
CONFIDENTIAL