Authorized licensed use limited to: Concordia University Library.

Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Software in Silicon in the Oracle SPARC M7
Processor
Kathirgamar Aingaran
Director, Hardware Development

Sumti Jairath
Sr. Hardware Architect

David Lutz
Principal Software Engineer

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. |

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Safe Harbor Statement
The following is intended to outline our general product direction. It is intended for
information purposes only, and may not be incorporated into any contract. It is not a
commitment to deliver any material, code, or functionality, and should not be relied upon
in making purchasing decisions. The development, release, and timing of any features or
functionality described for Oracle’s products remains at the sole discretion of Oracle.

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 3

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Agenda

1 Motivation For SW in Silicon

2 SW in Silicon Features in SPARC M7
3 HW – SW Stack Description
4 Oracle In Memory Database Performance Measurements
5 Performance Results For Use Cases Beyond Oracle Database
6 Conclusion

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 4

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Oracle SPARC M7
• 32 core, 4.1GHz SPARC processor
• 64MB L3 Cache
• 8 x DDR4 Ports, ~150GB/s
• Self Governed Cores for Power
Management
• SW in Silicon Feature Set
• Custom Hardware For Targeted Applications
• Security in Silicon
• SQL in Silicon
• Capacity in Silicon
• Same feature set on SPARC T7, S7

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 5

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Motivation For Software in Silicon
• Chip and Server Trends • Industry Trends
– Limitation on frequency scaling – Value shift: Transactions -> Analytics
– ILP and TLP scaling close to peak – Highly data parallel algorithms
– Power wall: – Faster persistent storage
• limit on power a system can deliver and • Focus on managing, mining, and securing
dissipate data
– Dark Silicon: – HW-SW Co-Design offers unique
• limit on simultaneous activity within chip opportunities
– Specialized, power efficient, hardware • FPGA, On Chip Accelerators, GP-GPU
provides advantage over general – Especially for full stack vendors like
purpose cores Oracle
=> SW in Silicon => SW in Silicon

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 6

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Agenda

1 Motivation For SW in Silicon

2 SW in Silicon Features in SPARC M7
3 HW – SW Stack Description
4 Oracle In Memory Database Performance Measurements
5 Performance Results For Use Cases Beyond Oracle Database
6 Conclusion

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 7

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Feature Candidates for Software in Silicon
• Analytics: Simple, data-parallel SQL primitives (marked with a *)

Processes: SQL:
Decode values*, Sum aggregation* select sum(lo_extendedprice*lo_discount) as revenue
Hash Joins from lineorder, date_dim
Bloom Filter Joins* where lo_orderdate = d_datekey and
Simple Filters* d_year = 2012 and
Range Filters* lo_quantity between 6 and 25 and lo_discount between 1 and 3

• Inline Decompression: “free” • Enhanced Security in Memory

decompression • Added layer of protection for multi-
• Reduces memory footprint required to tenant cloud environments
fit entire DB in DRAM/NVM

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 8

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Security In Silicon
• In-Memory Database
– Puts terabytes of data in the cloud
– Better be well protected
• Two types of Silicon enhanced protection
– Encryption in Silicon
• HW acceleration of AES, RSA, ECC, SHA-256/512, MD5,
Camellia, 3DES
– Silicon Secured Memory
• HW colors all data in memory
• Colors assigned by sys call
• Data can only be accessed with correct “color” in pointers
• Support for 14 colors (aka ADI states) at 64B granularity
Memory Silicon
Pointers Secured
Memory
Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 9
Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Silicon Secured Memory in Action
• Heartbleed style buffer overflow attack OLTP Workload
(normalized trans/hour)
2
Critical Routine
ld … 1.5
st …
hc28
P4ssw0rd# 1 SSM Disabled
SSM Enabled
Compromised 0.5
64Bytes ld … Routine
st … 0
150 users 250 users 350 users
SSM

• An additional layer on top of all existing security mechanisms. It is not a

replacement for encryption
• No performance loss for using Silicon Secured Memory

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 10

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
SQL in Silicon: Data Analytics Accelerator (DAX)

Local SRAM M7 DAX (1 of 8 Units)

Data Decompress Unpack/ Predicate Result Data

Align Eval Format/Encode Output
Input
Queues Decompress Unpack/ Predicate Result Queues
On-Chip
On-Chip Align Eval Format/Encode
Network Predicat Network
Data Decompress Unpack/ Result Data
Align e Eval Format/Encode
Input Output
Predicate Result
Queues Decompress Unpack/ Queues
Align Eval Format/Encode

• Stream Processor for data parallel operations

• DSP-like pipe for efficient filtering operations, typical of first phase of any query
• Cache sparing design with more complex processing in general purpose cores

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 11

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
DAX Features and Commands
• DAX executes vector commands on entire • Usage
streams of data – First step in query plan is usually a
• Scan: WHERE clause
e.g. SELECT CAR FROM C
– Does simple equality and range predicates
WHERE C.MAKE = “TOYOTA” …
• Translate: – Further trimming in join steps
– Applies Bloom Filter based predicates AND C.DEALERID = D.ID
AND D.LOC = “San Jose” …
• Select:
– More complicated steps in general
– Filters input streams, discarding unused data
purposes cores
• Extract: ORDER BY <distance from me>
– Converts memory formats to register friendly
formats
– E.g. Run length expansion (RLE), bit unpacking
Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 12
Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Capacity in Silicon
• For analytics applications and data warehouses
• Most data is nicely compressible
• Most data is immutable
• OZIP
• Special HW-SW co-designed compression algorithm in Oracle DB 12.1
• Reduces memory footprint by 2–5x
• Extremely fast: decompress at close to memory speed, 120GB/s
• Inline decompressors in DAX
• Decompress data in the path from memory to the general purpose cores.
• Or decompress inline when running DAX commands
E.g compressed scan runs filter directly on compressed data

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 13

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Agenda

1 Motivation For SW in Silicon

2 SW in Silicon Features in SPARC M7
3 HW – SW Stack Description
4 Oracle In Memory Database Performance Measurements
5 Performance Results For Use Cases Beyond Oracle Database
6 Conclusion

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 14

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
DAX Programming Model and Task Offload
• DAX calls are similar to system calls
Prepare Read • But fork a new “DAX thread” instead
Inputs (Core Thread) Results
of consuming core resources
initiate complete • Calling thread continues to execute while
(DAX thread)
DAX function proceeds asynchronously
• Parameters passed through shared
User Applications memory
• Results sent back through memory
Libdax API (thread safe) • Completion signaling through mwait()
OS Layer • Libdax API hides communication
Hypervisor HW Manager overhead in a simple function call
DAX • Architectural Invariant
• Easily called in Java, Scala, Python, C, …

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 15

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Integration with Oracle Database In-Memory
• Full stack integration with the DAX when In-Memory is enabled
• Automatic platform detection
– Best platform-specific libraries are chosen at run-time
– DAX enabled libraries are transparently used as appropriate
• Automatic acceleration of database queries
– No administrative changes required compared to other platforms
– No application changes required

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 16

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Oracle Database In-Memory Dual Format Database
• BOTH row and column
Memory Memory formats for same table
• Simultaneously active and
transactionally consistent
SALES SALES • Analytics & reporting use new
Row Column in-memory Column format
Format Format • New Analytics Compression means
huge amounts of database can now fit
in memory
SPARC
• Automatically detects and
uses Software in Silicon on M7
• OLTP uses proven row format
Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 17
Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Oracle Database In-Memory Column Store
• M7 Software in Silicon is transparently leveraged based on per-object MEMCOMPRESS
compression level and content analysis
• Encoding/compression options and heuristics are specific to each level
• All levels include Software in Silicon acceleration features
Compression Level Description
NO MEMCOMPRESS Data is populated without any compression
MEMCOMPRESS FOR DML Minimal compression optimized for Data Manipulation Language performance
MEMCOMPRESS FOR QUERY LOW Optimized for query performance (default)
SPARC M7 enables accelerated RLE processing
MEMCOMPRESS FOR QUERY HIGH Optimized for query performance as well as space savings
SPARC M7 enables OZIP compression with decompress/scan in one op
MEMCOMPRESS FOR CAPACITY LOW Balanced with a greater bias towards space saving
SPARC M7 enables OZIP compression with accelerated decompress
MEMCOMPRESS FOR CAPACITY HIGH Optimized for space saving

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 18

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Libdax Open APIs
• New APIs for SPARC M7/T7/S7 DAX features, available as libdax.so.1
– Available today on the Software in Silicon Developer Program Portal (see below), will be included in a
Solaris 11 support repository update (SRU) later in CY16
– Provides building blocks for analytics acceleration with any application
– post and poll commands, debug, log and dtrace
probe functions
– Includes OZIP library so apps can use HW
decompression features
• Software in Silicon Developer Portal
– https://swisdev.oracle.com
– Test drive Software in Silicon technology
– Learn and share

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 19

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Agenda

1 Motivation For SW in Silicon

2 SW in Silicon Features in SPARC M7
3 HW – SW Stack Description
4 Oracle In Memory Database Performance Measurements
5 Performance Results For Use Cases Beyond Oracle Database
6 Conclusion

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 20

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Timing Traces of Generic vs DAX Accelerated Libraries
Aggregate of 16 analytic queries based on Scale Factor 1000 Star Schema Benchmark

20 15
23x speedup of Scan 10 17x speedup of
10 processing Scan-Range processing
5
0 0
Scan Scan Range

10 1.5

10x speedup of Translate 1 30% speedup of Select

5
processing (a component 0.5
processing (a component
of Bloom Filter) of Data Projection)
0 0
Translate Select
• Includes multiple column widths and encodings with “memcompress for query low”
• For background on Star Schema Benchmark, see: http://www.tpc.org/tpctc/tpctc2009/tpctc2009-17.pdf
Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 21
Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Timing Traces of Generic vs DAX Accelerated Libraries
Decompression Acceleration
8
• Oracle In-Memory automatically
6
8x speedup of RLE chooses compression algorithm
4
Burst based on MEMCOMPRESS level,
2
0 performance heuristics, and data
RLE content
10 • Decompression is accelerated by
11x speedup of OZIP DAX hardware
5
Decompression • Compressed data can also be
0 scanned by the DAX without first
OZIP
decompressing
Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 22
Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Performance Gain for Oracle In-Memory
Generational comparison: SPARC M7 vs SPARC T5

• 4-socket SPARC M7 Logical Domain vs 4-socket SPARC T5-4

• Solaris 11.3, Oracle 12.1.0.2
• Workload based on Scale Factor 1000 Star Schema Benchmark
– 668 GB uncompressed on-disk
• All tables configured as “inmemory memcompress for capacity low”
– Enables compression/encodings including full OZIP on both platforms
• Concurrent analytic query streams scaled to 4 processes per core
• Key performance metric: Queries Per Hour (QPH)

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 23

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Performance Gain for Oracle In-Memory
Generational comparison: 4-socket SPARC M7 vs 4-socket SPARC T5
10 10
9x throughput gain 11x gain in performance
5 (Queries per Hour) with 5 per watt (Processor) due
M7 Software in Silicon to DAX efficiency
0 0
QPH QPH/watt

15 3 3x reduction in core
utilization due to DAX
10 14x gain in performance 2
offload – makes
relative to chip
5 1 resources available for
temperature
0 0 other business
QPH/°C %idle cores processing
• Concurrent analytic query workload with 4 PQ processes per core
• Workload based on Scale Factor 1000 Star Schema Benchmark
Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 24
Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Agenda

1 Motivation For SW in Silicon

2 SW in Silicon Features in SPARC M7
3 HW – SW Stack Description
4 Oracle In Memory Database Performance Measurements
5 Performance Results For Use Cases Beyond Oracle Database
6 Conclusion

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 25

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Analytics and Machine Learning (ML) on DAX
Examples of uses and performance

Area Analytics DAX Speedup

Machine Learning K-Nearest Neighbor 4x to 12x
Data Scanning Top N data in memory 4x to 7x
Data Scanning SQL on JSON data 4x to 5x
Streaming Tweet Analysis 5x to 9x

• SPARC DAX Offload acceleration with DAX API

– Easily used in Java, Scala, Python, C, C++, …
– Applicable to a wide variety of algorithms
– University research is also finding new creative uses of DAX API

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 26

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Quartet FS: In-memory Analytical Queries
Queries are 6x to 8x faster with DAX
• Higher query performance with lower core utilization Speedup on
– DAX offloads cores, 2x to 8x reduction in core utilization
8x queries
– Analytics: quartetfs.com
22%

20% Core Utilization 6x

18%
% core utilization

16%

14%

12%
DAX CPU Usage %
10%
No DAX CPU Usage %
8%

6%

4%

2%

0%
Streams 4 8 16 32 64 128 Sample 1 Sample 2

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 27

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Conclusion
1 SW in Silicon is custom hardware targeted at specific higher level
functions traditionally implemented in software
2 Cloud based applications, especially analytics, offer many opportunities
for SW in Silicon features
3 SW in Silicon in SPARC M7 provides gains in Performance, Power,
Security and Memory Capacity
4 Oracle database automatically uses these SW in Silicon features. Other
applications access through public API
5 Oracle is researching tighter HW-SW codesign, targeting deeper gains
on a wider set of applications
Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 28
Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Acknowledgments
• The authors wish to acknowledge the contributions of the SPARC M7 architecture,
design, firmware, and performance analysis teams, as well as the Oracle Database In-
Memory team, for bringing Software in Silicon to fruition, and thank the team members
for many helpful suggestions towards this presentation.
• The authors also wish to thank the many Oracle engineers who reviewed this slide set
and provided valuable feedback

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. | 29

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Software in Silicon in the Oracle
SPARC M7 Processor

Kathirgamar Aingaran
Sumti Jairath
David Lutz

Copyright © 2016, Oracle and/or its affiliates. All rights reserved. |

Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.
Authorized licensed use limited to: Concordia University Library. Downloaded on January 26,2022 at 15:37:47 UTC from IEEE Xplore. Restrictions apply.