Showing posts with label HotChips. Show all posts
Showing posts with label HotChips. Show all posts

Thursday, August 20, 2015

M7: Next Generation SPARC and Next Generation Data Center

[Oracle SPARC M7 Die, Courtesy The Register]

Abstract:

The SPARC Processor family from Sun Microsystems had existed for nearly 30 years. SPARC was an early contender in the 64 bit processing market while most commodity processors were still 32 bits or fewer. With the purchase of Sun Microsystems by Oracle, SPARC development continued and produced the fastest processor on the planet. Oracle promised a day where Oracle would have one processor for both the T and M platforms. The latest generation of processor finally appears able to unify the T and M system lines. In December 2014, the Hot Chips 26 Symposium in 2014 conference material was released to the general public, illustrating what is coming in 2015. Happy New Year!

[M7 32x Socket Interconnect, courtesy Enterprise Tech]

SPARC M7 at Hot Chips 26 in 2014:

Oracle announced during the Hot Chips conference that the new M7 processor and system would be released in 2015. Some notable reviews of the Oracle SPARC M7 included:
  • [2014年08月13日] Timothy Pricket Morgan from Enterprise Tech Systems Edition
    Oracle Cranks Up The Cores To 32 With Sparc M7 Chip
    When asked about what the performance advantage would be comparing an InfiniBand or Ethernet cluster running Oracle RAC and the Sparc M7-Bixby setup using the non-coherent memory clusters, Fowler said that the difference “would not be subtle.”
  • [2014年08月18日] Simon Sharwood from The Register
    Oracle reveals 32-core, 10 BEEELLION-transistor SPARC M7
    New chip scales to 1024 cores, 8192 threads 64 TB RAM, at speeds over 3.6GHz
  • [2014年08月12日] James Niccolai from PC World
    Oracle's Sparc M7 chip to supercharge in-memory computing
These reviews were helpful in understanding the market's take on the hardware announcement from Oracle, but the actual presentation was released in December 2014 had the details.

[Oracle 2010 SPARC Road Map, courtesy Enterprise Tech]

Hot Chips Oracle Presentation:

Stephen Phillips, the Senior Directory SPARC Architecture from Oracle, gave the presentation. He was involved in the architecture for T2+ (Victoria Falls) and later T-Series; M5, M6, M7 Processors, placing him in the cross-hairs of the delivery of the 2010 roadmap above. The Hot Chips 26 - August 12, 2014 (Big Iron Presentations) was included in Video (Video Start-End: 32:26-1:04:00) as well as High-Resolution PDF.

It should be noted, that the clustering of the M7 produces a system capable of remote memory sharing with up to 64 sockets, make the formerly released 5 year road map (pictured above) astounding accurate, illustrating the intense fidelity that Oracle offered to the SPARC and UNIX communities.

[M7 Decompression and Query Offload Engine, courtesy EnterpriseTech]

SPARC M7 In-Silicon Enhancements:

As one watches the presentation and reviews the high resolution slides, the following is a short set of notes, highlighting what some would consider key aspects.

Slide 3 - Recent history of SPARC
Slide 5 - M5 CPU - 20nm process; 32 Cores per Socket; 4 clustered cores; S4 enhanced core
Slide 6 - S4 Core - Dynamic Threading (1-8 threads) for speed & throughput; Faster Live migration
Slide 7 - Core Cluster - 1.5x Larger L2 Cache; 2x Greater Core Bandwidth
Slide 8 - Level 3 Cache & Network: 2.5x-5x Bandwidth; 25% Less Latency; HW Accelerator Access
Slide 9 - OVM "Aware" Solaris Process Groups by Core Cluster &, L3 Cache Partition
Slide 10 - Power aware in silicon; auto-adjusts voltage & frequency according to policy
Slide 11 - 2x-3x Memory Bandwidth; Live DIMM Retirement; Memory Lane Failover;
Greater than 2x PCIe performance
Slide 12 - Performance increase ~3x over M6
Slide 13 - Live Production Data Integrity Checking (for buffer overrun protection)
Slide 14 - Fine Grain Memory Migration for JAVA (for concurrent operations of middleware)
Slide 15 - Virtual Memory Masking for Java Runtime (embed object state into unused 64 bit)
Slide 16 - Decompression & Query accelerators for Oracle 12c (row & column for OLTP & OLAP)
Slide 17 - 8x Fused Decompression + Query Accelerators
Slide 18 - High performance for in-memory database without OS intervention
Slide 19 - 10-to-1 Decompression improvement of 1x query pipeline to 1x T5 (S3) Thread
Slide 20 - Third Party benefit through tool-chain
Slide 21 - Glue-less 1-8 socket support, like the T5
Slide 22 - SMP Scalability comprising 32 M7 SPARC Processors; 4 socket physical domains
Slide 23 - Reliable & Secure Shared Memory Clustering; 64 M7 sockets in a cluster; >1 failure tolerance
Slide 24 - Coherent Memory Cluster comprising 64 M7 SPARC Processors; secure foreign memory

The use of hardware application accelerators has proven to be a massive game changer in the industry for Oracle, as SPARC continues where others have failed.

Conclusions:

In the world of Network Management, this particular hardware solution enables massive network management scalability, the fastest virtualization migration technology, and provide the most reliable underlying infrastructure. Underpinning the SPARC processor technology includes 2x-10x "everything" over commodity Intel platforms, lashed together by various vendors. This is where high performance data centers which need a small footprint, low power utilization, enterprise software, and massive scalability will go.

Friday, March 29, 2013

Hot Chips 24: SPARC T5 Overview


[SPARC T5 Processor, courtesy Oracle 2013年03月26日 Announcement]
Abstract:
Every year, the best of engineering talent comes together in academia for Hot Chips conference, to present the best system designs. During Hot Chips 24, Session 9 - the SPARC T5 was presented by Sebastian Turullols and Ram Sivaramakrishnan from Oracle on Wednesday, August 29, 2012. This processor was released 6 months later, by Oracle with their T5 systems on Tuesday March 26, 2013.

Video Presentation:
The video presentation of Session 9 was conducted by Fujitsu, Oracle, and IBM. The middle portion of the presentation, starting at 30 minutes, includes the Oracle presenters.
[フレーム]

Slide Presentation:

The following screen shots were taken of the presentation. The full presentation is available here.




[4x memory controllers are capable of a peak of 128 GigaBytes per Second]











[Acceleration of Contended Locks; linked list of address requests; all requests satisfied atomically]

[Directory based level 3 indices, tracked up to 8 sockets on an on-chip SRAM]

[7 links with 14 lanes per link between sockets]


[C2C is a sharing cache]


[28 GigaBytes bandwidth between nodes; allows for aggregation of throughput via intermediate node]




[OLTP workload is extremely shared, makes workloads very difficult to scale]

[Elastic Mode includes all of the power saving features automatically]


[Solaris makes frequency request from hypervisor; cores reduced or cycle skipping used]



[Frequency and Voltage pushes performance maximum possible by the system]








Question-Answer Session: The following information was provided during the Question-Answer session, from the audience.
  • Voltage required for 3.6GHz varies from part to part. 0.95-1v is needed.
  • There is one voltage supply for all cores; one common PLL for the entire chip. Cycle skipping is used to vary cycle rate.
  • Low latency clustering port leverages "Allocated DMA Feature"
  • T5 no longer has an integrated 10 Gigabie Ethernet Controller.

Tuesday, August 10, 2010

Sorry Cliff Saran: UltraSPARC T3 Almost Here!


A Very Wrong Prediction


The Bizarre Prediction:
Cliff Saran, the managing editor of ComputerWeekly.com made a terribly bizarre prediction that no one in their right mind could ever consider as reasonable - that the "SPARC roadmap looked dim". How did he come to that conclusion?


The Facts Missed:
During the Hot Chips 2009 conference, there was a clear description of the architecture for the up and coming UltraSPARC 3 , at that point named Rainbow Falls. There was also a next generation crypto engines presentation. Clearly, had Cliff had not been watching the discussions from February 2010 about the 2 Billion Transistor UltraSPARC T3. Also missed was the code update in OpenSolaris for the official name UltraSPARC T3. Cliff apparently missed the additional code updates from OpenSolaris in July 2010 leveraging the official naming for the UltraSPARC T3.



The Freudian Slip:
Oracle held a web conference today (Tuesday August 10, 2010) talking about the SPARC roadmap, slipped [typo'ed?] a piece of information. Timothy Prickett Morgan covered the announcement from The Register (thanks for the screen shots!) indicating that the current generation of UltraSPARC has 512 threads (which is only possible on an UltraSPARC T3.) It seems like something was only partially redacted from the presentation, since the UltraSPARC T3 in a 4 sockets configuration should offer 64 cores with those 512 threads.


The Revelation:
Furthermore, John Fowler (formerly of Sun and now in Oracle) publicly released this rough image detailing the SPARC roadmap for the next 5 years.



The Conclusion:
For an industry managing editor to publish such a bizarre prediction in the headline of an article in his journal, discounting the information ringing in the development circles, and to be so wrong about his prediction just days before a formal announcement... indicates this editor is not an insider who has a tap on reality.
Subscribe to: Comments (Atom)

AltStyle によって変換されたページ (->オリジナル) /