Friday, December 3, 2010

Scalable Highest Performing Clusters at Value Pricing

Scalable Highest Performing Clusters at Value Pricing

Oracle presented another milestone achievement in their 5 year SPARC/Solaris road map with Fujitsu. John Fowler stated: "Hardware without Software is a Door-Stop, Solaris is the gateway."

The following is a listing of my notes from the two sessions. The notes have been combined, with Larry Ellison outlining the high-level and John Fowler presenting the lower-level details. SPARC T3 making world-record benchmarks. New T3 based integrated products. Oracle's Sun/Fujitsu M-Series gets a speed bump. SPARC T4 is on the way.

Presentation Notes:

New TpmC Database OLTP Performance
  • SPARC Top cluster performance
  • SPARC Top cluster price-performance
  • (turtle)
    HP Superdome Itanium 4 Million Transactions/Minute
  • (stallion)
    IBM POWER7 Power 780 10 Million Transactions/Minute
    (DB2 clustered through custom applications)
  • Uncomfortable 4 month for Oracle, when IBM broke the Oracle record
  • (cheetah)
    Sun SPARC 30 Million Transactions/Minute
    (standard off-the-shelf Oracle running RAC)
  • Oracle/Sun performance benchmark => ( IBM + HP ) x 2 !
  • Sun to IBM Comparison:
    3x OLTP Throughput, 27% better Price/Performance, 3.2x faster response time
  • Sun to HP Comparison:
    7.4x OLTP Throughput 66 Better Price/Performance, 24x compute density
  • Sun Supercluster:
    108 sockets, 13.5 TB Memory, Infiniband 40 Gigabit link, 246 Terabytes Flash, 1.7 Petabytes Storage, 1 Quadrillion rows, 43 Trillion transactions per day, 0.5 sec avg response

New Gold Release
  • Gold Standard Configurations are kept in the lab
  • What the customer has, the support organization will have assembled in the lab
  • Oracle, Sun, Cisco, IBM will all keep their releases and bug fixes in sync with releases

SPARC Exalogic Elastic Cloud
  • Designed to run Middleware
  • New T3 processor based
  • 100% Oracle Middleware is Pure Java
  • Tuned for Java and Oracle Fusion Middleware
  • Load-balances with elasticity
  • Ships Q1 2011
  • T3-1B SPARC Compute Blades based
    30 Compute Servers, 16 cores/server, 3.8 TB RAM, 960 GB mirrored flash disks, 40 TB SAS Storage, 4 TB Read Cache, 72 GB Write Cache, 40 Gg/sec Infiniband, 10 GigE to Datacenter

SPARC Supercluster
  • New T3 processor based and M processor based
  • T3-2 = 2 nodes, 4 CPU's, 64 cores/512 threads, 0.5 TB RAM, 96 TB HDD ZFS, 1.7TB Write Flash, 4TB Read Flash, 40 Gbit Infiniband
  • T3-4 = 3 nodes, 12 CPU's, 192 cores/1536 threads, 1.5 TB RAM, 144 TB HDD ZFS, 1.7TB Write Flash, 4TB Read Flash, 40 Gbit Infiniband
  • M5000 = 2 nodes, 16 CPU's, 64 core/128 threads, 1 TB RAM, 144 TB HDD ZFS, 1.7TB Write Flash, 4TB Read Flash, 40 Gbit Infiniband

T3 Processor in production
  • Releases already, performing in these platforms
  • 1-4 processors in a platform
  • 16 cores/socket, 8 threads/core
  • 16 crypto-engines/socket
  • More cores, threads, 10 GigE on-chip, more crypto engines

T4 Processor in the lab!
  • Thousands under test in the lab, today
  • To be released next year
  • 1-4 processors
  • 8 cores/socket, 8 threads/core
  • faster per-thread execution

M3 Processor from Fujitsu
  • 1-64 SPARC64 VII+ Processors
  • 4 cores, 2 threads/core
  • Increased CPU frequency
  • Double cache memory
  • 2.4x performance of original SPARC64 VI processor
  • VII+ boards will slot into the VI and VII board chassis
Flash Optimization
- Memory hierarchy with software awareness

- Appropriate for High Performance Computing
- Dramatically better performance than Ethernet for linking servers to servers & storage

New Solaris 11 Release

  • Next Generation Networking
    re-engineered network stack
    low latency high bandwidth protocols
  • Cores and Threads Scale
    Adaptive Thread and Memory Placement
    10,000's of core & threads
    thread observability with DTrace

  • Memory Scale
    Dynamic optimization for large memory configs
    Advanced memory placement
    VM systems for 1000's TB memory configs

  • I/O Performance
    Enhanced NUMA I/O framework
    Auto-Discovery of NUMA architecture
    I/O resources co-located with CPU for scale/performance

  • Data Scale
    ZFS Massive storage for massive datasets

  • Availability
    Boot times in seconds
    Minimized OS Install
    Risk-Free Updates with lightweight boot and robust package dependency
    Extensive Fault Management with Offline failing components
    Application Service Managemment with Restart failed applications and associated services quickly

  • Security
    Secure by default
    Secure boot validated with onboard Trusted Platform Module
    Role Based Root Access
    Encrypted ZFS datasets
    Accelerated Encryption with hardware encryption support

  • Trusted Solaris Extensions
    Dataset labels for explicit access rules
    IP labels for secure communication

  • Virtualization
    Network Virtualization to add to Server and Storage Virtualization
    Network Virtualization includes Virtual NIC's and Virtual Switches
SPARC Supercluster Architecture
  • Infiniband is 5x-8x faster than most common Enterprise interconnects
    Infiniband has been leveraged with storage and clustering in software
  • Flash is faster than Rotating Media
    Integrated into the Memory AND Storage Hierarchy

SPARC 5 Year Roadmap
  • SPARC T3 delvered in 2010
  • SPARC VII+ delivered in 2010
  • Solaris 11 and SPARC T4 to be delivered in 2011
Next generation of mission critical enterprise computing
  • Engineer software with hardware products
  • Deliver clusters for general purpose computing
  • Enormous levels of scale
  • Built in virtualization
  • Built in Security
  • Built in management tools
  • Very Very high availability
  • Tested with Oracle software
  • Supported with Gold Level standard
  • Customers spend less time integrating and start delivering services on systems engineered with highest performance components

No comments:

Post a Comment