Performance of OpenVMS
v8.4 with i2 servers
Rafiq Ahamed K,
OpenVMS Engineering
©2011 Hewlett-Packard Development Company, L.P.
©2010
The information contained herein is subject to change without notice
Agenda
•
What's new in i2 Servers?
•
Introduction
– New
i2 server a performance differentiator
– Performance
– Future
•
Enhancements
Focus
Performance Results
– Platform
– Application
– Alpha
to i2 server Migration Differences
•
Summary
•
Q&A
2
9/19/2011
―The performance results shared in this session are from
engineering test environment, they do not represent any
specific customer workload. Your mileage may vary.‖
3
9/19/2011
What is new in i2 servers,
post i2 blades?
i2 server family
Introduced rx2800 i2 rack-mounted server
OpenVMS V8.4
Integrity 2 Socket Rack Server
Integrity Server Blades
World’s first scale-up
blades built on
the industry’s #1
blade infrastructure
New!
5
9/19/2011
8-core scalability with 3x improved
density
without sacrificing RAS
Introducing rx2800;
a performance differentiator
rx2800 i2 Server Architecture
7
9/19/2011
CPU-CPU 19.2 GB/s
Memory: 28.8GB/s peak per Processor Module
QPI, IOH to Processors: 38.4 GB/s
Intel 9300 Processor
A performance differentiator
8
9/19/2011
Performance Features of i2 Servers
Memory
• Dual Integrated Memory
Controllers with 4 SMI channels,
peak memory band-width up to 34
GB/s (6x)
MB
MB
MB
• Intel® Scalable Memory Interconnect
(Intel® SMI), connects to the Intel®
7500 Scalable Memory Buffer (SMB) to
support larger physical memory DDR3
RDIMMs
MB
SMI
MB
• DDR3 Higher Throughput
(800MT/s), Lower Power, Faster
Response Time, Increased
Capacity/DIMM (16GB)
MB
• Directory-based Cache
Coherency – Reduces Snoop
traffic and contention
DDR3 DIMM DDR3 DIMM DDR3 DIMM DDR3 DIMM
DDR3 DIMM DDR3 DIMM DDR3 DIMM DDR3 DIMM
DDR3 DIMM DDR3 DIMM DDR3 DIMM DDR3 DIMM
DDR3 DIMM DDR3 DIMM DDR3 DIMM DDR3 DIMM
DDR3 DIMM DDR3 DIMM DDR3 DIMM DDR3 DIMM
DDR3 DIMM DDR3 DIMM DDR3 DIMM DDR3 DIMM
MB
• 1MB Directory Cache/IMC
Processor
MB
• Capability to supports up to 1TB
memory per IMC
• 6 Gen 2 PCIe slots, supporting 5GT/sec
• p410i SAS 6G RAID Controller
• 8 10K RPM Disks
• Enhanced Thread-Level
Parallelism (TLP) [8T/P]
• Instructions-level parallelism
(ILP) – minimize threads from
stalling the pipeline
• Intel® Turbo Boost Technology
– Performance on Demand
• Intel® VT-i2 is Introduced
Itanium® 9300
(Tukwila-MC)
Itanium® 9300
(Tukwila-MC)
QPI
Intel® 7500
IOH
(Boxboro-MC)
• Data TLB support for 8K and
16K pages
QPI
Intel®
ICH10
• New Intel® QuickPath
Interconnect Technology replaces the Front Side Bus
with a point-to-point
• 4 full-width Intel QuickPath
Interconnect links and 2 halfwidth links per processor
• SMB supports different size and types
of DIMM
IO
NUMA
Aware!
PCIe Gen2
Gen1
PCIe Devices
PCIe Devices
• Peak processor-to-processor
and processor-to-I/O
communications up to 96 GB/s
(9x)
• Glueless System Designs Up to
Eight Sockets – FSB
Limitations
Key Characteristics
Intel® Itanium® processor 9100
Intel® Itanium® processor 9300
Cores
2
4
Total On-Die Cache
27.5 MB
30 MB
Software Threads per Core
2
2 (w/ enhanced thread management)
System Interconnect
(bandwidth per processor for
a 2-socket system)
Front Side Bus
• Peak bandwidth per processor: 5
GB/s
Intel® QuickPath Interconnect Technology
• Peak bandwidth: 48 GB/s (up to 9x
improvement)
• Enhanced RAS
• Enables common IOHs with nextgeneration Intel® Xeon® processors
Memory Interconnect
(bandwidth per processor for
a 2-socket system)
Front Side Bus
• Peak bandwidth per processor: 5
GB/s
Dual Integrated Memory Controllers
• Peak bandwidth 34 GB/s (up to 6x
improvement)
Memory Capacity
(4-socket system)
128-384 GB
1TB (using 16 GB RDIMMs) — up to 8x
improvement
Partitioning and
Virtualization
Intel® VT-i
Intel® VT-i2
Energy Efficiency
Demand Based Switching (DBS)
•
•
•
SMP Scalability
10
9/19/2011
Source: Intel Corporation
•
•
•
64-bit Virtual Addressability
50-bit Physical Addressability
Home snoop coherency
•
•
•
•
Enhanced DBS (voltage modulation in
addition to frequency)
Intel® Turbo Boost Technology
Advanced CPU and Memory Thermal
Management
64-bit Virtual Addressability
50-bit Physical Addressability
Directory coherency for better
performance in large SMP configurations
Up to 8-socket Glueless systems (higher
scalability with OEM chipsets)
OpenVMS V8.4 Performance
Enhancements
(includes post V8.4)
11
V8.4 Major Performance Features
•
Resource Affinity Domain (RAD) support for IA64
•
Packet Processing Engine (PPE) Support for TCP/IP
•
Automatic Dynamic Processor Resilience (DPR)
•
Compression support for BACKUP
•
RMS SET MBC count support for 255 blocks
•
XFC support for 4G Global Hint Pages
•
Shadowing T4 Collectors
•
File System Consistency Check Control for reducing directory corruption
•
Asynchronous Virtual IO (AVIO) support for Guest OpenVMS running on
HPVM
V8.4 Performance Enhancements..
•
Core OS improvements
–Dedicated
–Exception
–Reduced
Lock Manager and Cluster Driver Optimizations
Handling Optimizations
Spin Lock Contentions (SCHED, MMG etc)
–Optimizations
in Global Section Deletion and Creation
Algorithms
–Introduced
–SYSMAN
–RTL
Paged Pool Look Aside List (LAL)
IO AUTO Performance Improvements (Fibre Only)
Changes to optimize strcmp() and memcmp()
–Support
for new high speed USB connectivity and enhancements
–rx2800
console performance fixed!
V8.4 Performance Enhancements..
•
Shadow feature improvements to WriteBitMap, MiniCopy,
MiniMerge and SPLIT_READ_LBN
•
Compiler improvements (BASIC RTL, COBOL, OTS$STRCMP)
•
Miscellaneous Improvements
– FLT
Tracing has 2 new options for SUMMARY report (using INDEX/PID)
– LIBRARIAN,
14
9/19/2011
MONITOR , DCL DELETE
OpenVMS 8.4 Performance Results
OpenVMS V8.4
Summary of Improvements
Significant performance gains across various workloads
and system configurations
Reduced BACKUP Compressed Data
272
Faster Shadow READ
300
Faster Bitmap Updating in SHADOWING
400
High Speed USB Boot
25
High Usage of mem(str)cmp()
350
Larger AST Queuing
50
Stressful DLM Resources
200
Heavy Exception Handling
50
Heavy Image Activation
45
0
100
% Improvement
200
300
400
500
Future Focus
•
Continuous work to optimize performance
on new servers with new processors
•
On going effort to reduce alignment faults
•
Maximize utilization using hyper
threading
•
Additional T4 collectors
•
Spin Lock Optimizations
•
Many more…
16
9/19/2011
Performance of i2 Servers;
Platform and Application Tests
Price/Performance Migration Path
New
Blades
NewIntegrity
Integrity blades:
Cores
8-Socket
16-core
rx7640
4-Socket
BL890c i2
8-core
rx6600
BL870c i2
2-Socket
rx3600
BL870c
2-Socket
rx2660
Blue is Price, Light Blue is Performance.
BL860c i2
rx2800
BL860c
New Integrity Blades are more
scalable than previous generation
• Double the number of cores
• More memory
• More I/Os
OpenVMS running on i2 servers
Performance Highlights
i2 servers architected for high performance
– Architecture
provides increased number and faster cores/socket
– Superior
memory and interconnect technology
– Memory
intensive applications benefit from low latency and high bandwidth architecture
– Higher
– More
time
IO bandwidth and throughput resulting from new IO architecture
headroom for CPU, Memory and IO intensive workloads with improved response
Upto 3x performance improvement with i2 servers running OpenVMS
– Our
test have shown up to 2x improvement with java, some database and web server
applications
– Oracle
19
9/19/2011
has shown 3x improvement
CPU Ratings

Integer Tests
Per
Processor
1.73 GHz/ 1.6 GHz 9300
series processor show 2-2.3x
performance improvement
Ratings
(More is Better)
4000
3500
3000
9300 - BL8x0c-i2 (1.73GHz/6.0MB)
2500
9300- BL8x0c-i2 (1.60GHz/6.0MB)
2000
9300 - BL8x0c-i2 (1.33GHz/4.0MB)
1500
9000 - BL860c (1.59GHz/9.0MB)
1000
9100 - rx7640 (1.60GHz/12.0MB)
500
0
– These numbers are per processor/socket
– As the frequency increases, we see a increase in rating
Applications with large Integer computations, financial analysis, high data processing
20
9/19/2011
CPU Ratings

Floating Point Computation Tests
Per
Processor
1.73 GHz/ 1.6 GHz 9300
series processor show 2.12.3x performance
improvement
FP Rating
(More is Better)
9000
8000
7000
6000
5000
4000
3000
9300 - BL8x0c-i2 (1.73GHz/6.0MB)
9300 - BL8x0c-i2 (1.60GHz/6.0MB)
9300 - BL8x0c-i2 (1.33GHz/4.0MB)
9000 - BL860c (1.59GHz/9.0MB)
9100 - rx7640 (1.60GHz/12.0MB)
2000
1000
0
– Intel ® Itanium ® 9300-series processors have new high precision
floating architecture
Fast response to complex operations; Scientific, Automation and robotic
applications should benefit
21
9/19/2011
Memory Tests
55% improvement
in memory
bandwidth
between 3600
and BL860c i2
Throughput – More is Better
Memory Bandwidth
Single
Stream
(More is Better)
3500
3000
2500
BL860c-i2 (1.73GHz/6.0MB)
2000
rx3600 (1.67GHz/9.0MB)
rx6600 (1.59GHz/12.0MB)
1500
rx7640 (1.60GHz/12.0MB)
1000
SD32A (1.60GHz/9.0MB)
500
0
MB/Sec
– Computed via Memory Test Program – Single Stream
– The new BL8x0c i2 server demonstrated very good memory bandwidth
– Memory bound applications should see a difference with aggregated
bandwidth
22
9/19/2011
Core I/O on rx2800
rx2800 i2 - Core SAS Caching
rx2800 i2 - SAS Logical Disk (Striping)
700
2500
600
2000
400
IOPS
IOPS
500
300
200
1500
1000
500
100
0
0
1
2
4
8
16
32
64
128
1
256
2
•
8
16
32
64
128
Cache
1 disk with Cache
2 disk with Cache
4 disk with Cache
rx2800 i2 server comes with p410i SAS Controller as Core SAS I/O
– Small
Block Random Tests were run on same sized disk
– Logical
Volumes were spread across multiple disks to show the I/O striping effect
•
p410i with cache* exponentially boosts the performance (upto 2.5x)
•
Increased number of disks in a raid group linearly increases performance
23
9/19/2011
256
Load
Load
W/O Cache
4
* Additional Cache Kit
Rack mounted - Core I/O i2 comparison
Integirty Core SAS
700
600
IOPS
500
400
300
200
100
0
1
2
4
8
16
32
64
128
256
Load
rx6600 (1.59GHz/12.0MB)
rx2800 i2 (1.46GHz/5.0MB) - Core Cache
rx2800 i2 (1.46GHz/5.0MB) - W/O Cache
•
Core IO SAS Controller (rx6600 - LSISAS106x, rx2800 – p410i)
•
Older IA64 SAS performance is similar to W/O Cache on i2 server
•
rx2800 i2 server with Cache is a clear winner ( upto 2.5x times better)
24
9/19/2011
Apache Performance
2x
Improvement
Apache Bench Tests on OpenVMS 8.4
600
Bandwidth
Throughput
Time Taken (sec)
(More is better)
( More is Better)
(Less is Better)
1400
531
1200
500
60
1162
1000
400
276.21
300
800
600
200
40
597
0
Transfer rate (Kbytes/sec)
25.818
10
200
0
30
20
400
100
50.277
50
0
Requests per second
Time Taken (sec)
BL860c i2 (1.73GHz/6.0MB)
BL860c i2 (1.73GHz/6.0MB)
BL860c i2 (1.73GHz/6.0MB)
BL860c (1.59GHz/9.0MB)
BL860c (1.59GHz/9.0MB)
BL860c (1.59GHz/9.0MB)
– Configuration Details
•
The tests were run on OpenVMS 8.4, with 1G network end to end connectivity
•
Apache 2.1-1 with ECO2 , Apache Bench 2.0.40-dev
– Time Taken should be less; Req/sec and KB/sec should be more
•
25
BL860c-i2 was able to cater 2x performance compared to BL860c
9/19/2011
Java Workload Tests
2x
Improvement
Native Java Tests on OpenVMS 8.4 – More is Better
Java Workload
140000
140000
120000
120000
Operation Rate
Operation Rate
Java Workload
100000
80000
60000
40000
100000
80000
60000
40000
20000
20000
0
0
0
2
4
6
8
10
12
14
16
18
8
9
10
BL870c i2 (1.60GHz/5.0MB)
rx6600 (1.59GHz/12.0MB)
– Java Workloads scale up better on i2 Servers
– Java Workloads are high CPU and Memory Intensive
26
9/19/2011
12
13
14
15
Threads
Threads
rx6600 (1.59GHz/12.0MB)
11
BL870c i2 (1.60GHz/5.0MB)
16
Rdb Tests
2x
Improvement
Rdb Performance Load Tests – OpenVMS 8.4
RDB Performance
RDB Performance
(Less is better)
250
1.2
200
1
150
sec/txn
TPS
(More is better)
100
0.8
0.6
0.4
50
0.2
0
0
HP Integrity rx2800 i2 (1.60GHz/5.0MB)
HP Integrity rx2800 i2 (1.60GHz/5.0MB)
HP rx3600 (1.59GHz/9.0MB)
HP rx3600 (1.59GHz/9.0MB)
– Test runs a load with100 jobs of 100000 txns/job, DB on local disks
(single)
– Rx2800 i2 server shows 2x better throughput than rx3600
•
•
The Transaction Per Second (TPS) is 2x
Rdb is high Memory and I/O oriented application, which seems to take advantage of p410i
6G SAS controller and high speed memory architecture
– Results have shown that Rdb takes advantage of HT in high end i2 servers
27
9/19/2011
Oracle 10gR2 on new i2 Server
3x
Improvement
20000
TPM
15000
10000
5000
0
16
32
48
Users
rx7640 (1.60GHz/12.0MB)
BL890c i2 (1.60GHz/6.0MB)
– Oracle Swing Bench Tests were run with tuning configuration (same)
• rx7640 and BL890c i2 are NUMA based systems, MostlyNUMA RAD
Enabled and Hyper-Thread disabled, 6 RAID 5 EVA8100 Volumes
– Oracle was run in Shared Server Mode
– BL890c i2 server consistently shows 3x improvement for same numbers
of users
28
9/19/2011
Oracle Tests – Resource Usage
16000
14000
12000
10000
8000
3x Increase
6000
4000
2000
0
CPU
TPM
BL890c i2 (1.60GHz/6.0MB)
rx7640 (1.60GHz/12.0MB)
– BL890c i2 is able to drive 3x improvement for same CPU usage
29
9/19/2011
ALPHA TO IA64
Speeds and Feeds
©2010 Hewlett-Packard Development Company, L.P.
The information contained herein is subject to change without notice
Mission-critical scaling
Speeds and Feeds comparison
GS1280 M32
BL890c i2
Processor Type
1150 MHz EV7
8 Intel® Itanium® processor
9300 series
(quad-core and dual-core)
Max Memory
256 GB
1.5 TB
Internal Storage
Ultra3 SCSI
6G SAS HW RAID
Networking
(integrated)
1 GbE NICs
10 GbE (Flex-10) NICs
PCI/PCI-X slots @133 MHz
PCIe Gen2 Mezz slots
2x 41U
4U
IO Slots
Form Factor
BL890c i2 scales to support mission-critical workloads
32
9/19/2011
Compilation Tests
Compilation Tests
(Less is better)
600000
557110
500000
400000
300000
242455
200000
100000
0
BL890C - 8S
Itanium 9300 (1.6GHz)
Alpha GS1280
EV7(1150 Mhz)
Charged CPU
•
This test is to compile around 200 modules using c++ compiler
•
The overhead of running this on BL890C i2 was reduced ~2.3 times
compared to GS1280
33
9/19/2011
Java Workload Tests
Operation Rate
Java Tests
200000
180000
160000
140000
120000
100000
80000
60000
40000
20000
0
1
5
9
13 17 21 25 29 33 37 41 45 49 53 57 61
Threads
BL890c i2
GS1280
– Java Workloads scale exponentially with load as compared to Alpha
–There is no active development involved in Java on Alpha
–Last available on alpha is JDK 1.5 as compared to JDK 1.6
– Java Workloads are high CPU and Memory Intensive
34
9/19/2011
EV7 vs. Itanium 9300(Tukwila)
Per Core
– Tukwila Processors are ~2.6
times faster than EV7
– These numbers are per Core
(within a processor/socket)
– Tukwila processors have new
high precision floating
architecture
Floating Point Ratings
(Higher is better)
2000
1882
1800
1600
1400
1200
1000
800
699
600
– Fast response to complex
operations; Scientific,
Automation and robotic
applications should benefit
35
9/19/2011
400
200
0
BL890C - 8S
Itanium 9300 (1.6GHz)
Alpha GS1280
EV7(1150 Mhz)
Floating Point Rate
EV7 vs. Intel Itanium 9300(Tukwila)
Per Core
– Tukwila Processors are ~33%
better than EV7
– These numbers are per Core
(within a processor/socket)
– CPU Bound applications should
benefit (database queries),
specifically integer
computational bound
applications
Integer Ratings
(Higher is better)
900
851.46
800
700
563.38
600
500
400
300
200
100
0
BL890C - 8S
Itanium 9300 (1.6GHz)
Integer Rate
36
9/19/2011
Alpha GS1280
EV7(1150 Mhz)
Memory Bandwidth Tests
•
•
This tests simulates heavy load
directly to memory module by
passing the cache
The bandwidth on a ―single
stream‖ test for BL890c i2 is
52% better
Memory Bandwidth
(Higher is better)
3500
3200
3000
2500
2100
2000
1500
1000
500
0
BL890C - 8S
Itanium 9300 (1.6GHz)
Memory (MB/sec)
37
9/19/2011
Alpha GS1280
EV7(1150 Mhz)
IO Tests
Single
Process
Tests
IO Tests on EVA8100 with 2/4Gb Cards, RAID 5 320GB Volumes
Random Workloads
Small Block Random Mixed Load
647.9
700
14000
12000
IOPS
500
400
325
287.2
300
200
100
112.7
Operation Rate
600
10000
8000
6000
4000
2000
0
0
1
Threads
AlphaServer ES45 Model 2B
2
4
8
16
32
64
128
256
Load
BL860c i2
AlphaServer ES45 Model 2B
BL860c i2
–4K IO Tests run with Mixed QIO/FastIO Loads, RANDOM workloads are
very CPU intensive
–We see ~3x increase in throughput for same load on i2 server
– On alpha platforms, the max device supported is 2G FC HBA’s and the bus speed and
feed influenced a lot in the performance
38
9/19/2011
Rdb Tests
RDB Performance
(More is better)
250
214
200
145
TPS
150
100
50
0
BL890C - 8S
Itanium 9300 (1.6GHz)
Alpha GS1280
EV7(1150 Mhz)
– Test runs a load with 100 jobs of 100000 txns/job, DB on local
disks (single)
– BL890c 8S performs approx.1.5x better than GS1280 (Alpha)
•
39
The Transaction Per Second (TPS) is 1.5x
9/19/2011
Performance Enhanced with the Integrity Server
Blades Based on Blade Scale Architecture
Up to
2X1
faster performance
Integrity server blades
based on Blade Scale
Architecture
Dual-core
Integrity servers
2- & 4-socket Integrity
with built-in
resiliency and less
power
consumption
1 some
Per socket performance increases
2.3x
Up to 2x
applications have shown up to 3x
Integer & Floating Tests
Application Performance
Summary
•
i2 servers replace the Front Side Bus (FSB) with Quick Path
Interconnect(QPI) Fabric
•
Even the low end i2 server comes with scalable features and supports
NUMA
•
The performance improvements delivered with i2 server are a mix of
– Hardware-level
– Operating
improvements
system innovations
•
Customers will see up to 2x performance benefits with i2 servers
•
The i2 servers support state of the art RAS capabilities
•
The TCO paper outlines the cost, power and density benefits
•
The speed and feed of i2 servers are very large compared to Alpha!
43
9/19/2011
References
•
Please mail us your OpenVMS performance
feedback or concerns to t4@hp.com
•
Please refer to below 2 white papers on
OpenVMS performance and scalability
– Performance
– Total
•
Benefits on BL8x0c i2 Server Blades
Cost of Upgrade
i2 Server Technical Reads
– Take
a tour around the new HP Integrity rx2800 server
(Video, 2:44)
– new
HP Integrity rx2800 technical white paper
– White
paper: Why scalable blades: HP Integrity server
blades
•
Didn’t find what you are looking for
– Enterprise
servers, workstations, and systems hardware
technical web site
44
9/19/2011
Questions/Comments
•
Business Manager (Rohini Madhavan)
– rohini.madhavan@hp.com
•
Office of Customer Programs
– OpenVMS.Programs@hp.com
THANK YOU
Boxboro & ICH10
•
Boxboro I/O Hub (E7500 Chipset)
Connects to local CPUs via QPI links
– Provides 36 PCIe Gen 2 lanes
–
•
–
Hosts major IO functions
•
•
•
•
•
•
Order of magnitude peak IO
bandwidth increase over previous
generation in BL870c i2
p410i RAID/SAS controller
Two dual-port 10GbE Flex-10 NICs
Three x8-provisioned mezzanine slots
Gromit XE (iLO3) mgmt controller
ICH10 I/O Controller Hub
(SouthBridge)
BIOH
2.5 GB/s
5 GB/s
5 GB/s
10 GB/s
10 GB/s
10 GB/s
ICH10 utilization
x4 PCIe Gen 1 link for partner blade
support
– Support VGA controller, USB
controller
–
53
1.2 GB/s
9/19/2011
Gromit XE
RAID/SAS Controller
Dual 10 GbE
Dual 10 GbE
Mezz Slot 1
Mezz Slot 2
Mezz Slot 3
ICH10 on ICH Card
2.5 GB/s
Peak bidirectional link bandwidth
Intel ® Itanium ® 9300 Changes
High Performance
Enhanced Thread-Level Parallelism (TLP)
•4 Threads/Processor; Improve performance and scalability for heavily threaded software
•Improved Thread management providing high core utilization; Thread switch for medium latency
events and spin locks
•Dedicated cache instead of shared cache model; Quick response, no contention on cache
Enhanced Instruction-Level Parallelism (ILP)
•Simultaneously process multiple instructions on each software thread; Increases throughput and faster
response
•Wide and short pipelining with many optimizations. Supports many zero cycle load and re-steering,
with extensive bypassing for not stalling the threads.
•1MB First-level Data Translation Lookaside Buffer (DTLB)/ Core supports larger pages (8 K and 16 K)
Intel® Turbo Boost Technology – Performance on Demand
•Automatically allows processor cores to run faster than the base operating frequency
•When workload demands additional performance, the processor frequency will dynamically increase
by 133 MHz on short and regular intervals (6microsec), 120 events monitored
Intel® VT-i2 Introduced - Provide additional performance optimizations
•HW assistance helping reduction in emulation code and OS-VMM transitions
54
9/19/2011
Changes/Improvements
Scalability and Headroom
Enhanced Communication Channels (Memory, IO) – Highly Scalable
• New Intel® QuickPath Interconnect Technology - replaces the Front Side Bus with a pointto-point
• 4 full-width Intel QuickPath Interconnect links and 2 half-width links
• Peak processor-to-processor and processor-to-I/O communications up to 96 GB/s (9x)
• Supports PCI Express Gen2 (5.0 GT/s) as well as PCI Express Gen1 (2.5 GT/s).
Two Integrated Memory Controllers and an Intel® Scalable Memory Interconnect
• 2 integrated memory controllers, peak memory band-width up to 34 GB/s (6x)
• Capability to supports up to 1TB memory per IMC
• Intel® Scalable Memory Interconnect (Intel® SMI), connects to the Intel® 7500 Scalable
Memory Buffer to support larger physical memory DDR3 RDIMMs (faster)
Directory-based Cache Coherency – Reduces Snoop traffic and contention
• Cache coherency ensures that data in memory and cache remain synchronized, so the
most current data is used in every transaction
• In combination with QPI provides excellent scaling
Glueless System Designs Up to Eight Sockets – FSB limitation of max 4 is eliminated.
55
9/19/2011
Memory tests
Memory Latency on NUMA Systems
Memory Latency
(Less is Better)
Latency(nsec)
Single
Stream
400
350
300
250
200
150
100
50
0
Local
rx7640 (1.60GHz/12.0MB)
56
Remote
Interleaved
BL860c-i2 (1.73GHz/6.0MB)
•
Computed via Memory Test Program – Single Stream
•
The local is 10% better, remote 28% and interleaved 20%
9/19/2011