Martin McKendry - Open Parallel, USA

Martin McKendry - Open Parallel, USA

“The Contribution of MultiCore to Server Architecture: Part II -Implications for the Future”

Bd1c4acb24d143c7ca8dff849461ebe3?s=128

Multicore World

July 16, 2012
Tweet

Transcript

  1. The Contribution of MultiCore to Server Architecture Part 2: Implications

    for the Future 1 Martin McKendry Steve Friedl
  2. What I do • Work in Silicon Valley – Best

    place to work in the world • Work at the intersection of market capitalization and technology – Growth companies: FileNet and Siebel – Value plays: Avaya and Openwave • Constantly figuring out what building blocks to use, how and where to get teams to use them 2
  3. Outline • Resources today: MultiCore Everywhere • Application Architecture •

    The World According to SQL • Breaking Assumptions – NoSQL – Flash • New Application Architecture • Startups Exploiting Technology Trends 3
  4. The New Age • We have plenty of hardware resources

    now – MultiCore Processors, Networking, Flash Storage • But we are restricted by legacy interfaces – Virtualization – Block Device Protocols – SQL 4
  5. On the eve of the release of Ivy Bridge, Intel

    is finally bringing its server chips up to speed by introducing theSandy Bridge-based E5-2600 family of CPUs. The company claims its latest processors outperform the previous generation of Xeons by up to 80 percent in raw speed, while improving per-watt performance by 50 percent. A grand total of 17 different Xeons will be available, ranging in price from $198 to $2,050. The eight-core chips support up to 768GB of RAM, PCI Express 3.0, Hyper-Threading, Turbo Boost, Intel Virtualization -- basically the whole Chipzilla portfolio of tricks. We have plenty of CPU power and addressable memory 5
  6. Specification FlashArray Controller (each) CPU 2x Intel Xeon 6-core CPUs

    RAM 48GB (working cache) BOOT DRIVES Redundant hot-swappable boot SSDs FRONT-END PORTS - 4x 8Gb/s Fibre Channel (SFP) - 1x empty expansion slot BACK-END PORTS - 4x 6Gb/s SAS Storage Shelf Interconnect - 2x 40Gb/s QDR Infiniband Controller Interconnect MANAGEABILITY - 2x GbE (RJ-45) - Serial and USB KVM access POWER SUPPLY Redundant, hot-swappable power supp Storage uses serious processor power 6
  7. Specification FA-320 (2 controllers, 2 storage shelves) FA-310 (1 controller,

    1 storage shelf) IOPS (TYPICAL 4K RANDOM) 300,000 200,000 SUSTAINED WRITE IOPS 180,000 140,000 BANDWIDTH 3 GB/sec 2 GB/sec SUSTAINED WRITE BANDWIDTH 1GB/sec 500 MB/sec LATENCY < 1 ms average latency < 1 ms average latency EFFECTIVE CAPACITY* - AT 5-TO-1 DATA REDUCTION - AT 10-TO-1 DATA REDUCTION Up to 100 TB Up to 200 TB Up to 50 TB Up to 100 TB To produce serious throughput 7
  8. Cisco Consolidates 40 customized packet-processor cores (900 MHz to 1.2

    GHz) into a single piece of silicon. Plenty of processing at the network interface 8
  9. We have the processor power, but… • Application software architectures

    are based on – Lack of CPU power – Lack of memory – Large distance to secondary storage – Speed of rotating media 9
  10. Application Architecture Permanent Storage Media (Disk) Memory CPU SQL Database

    Middleware Applications: HR, Finance, Sales, Service… 10 SQ L Block Device
  11. Oracle as Example • 2003 Buys JD Edwards $1.8B •

    2005 buys PeopleSoft $10.3B • 2005 Buys Siebel $5.8 B • 2008 buys BEA $8.5B – BEA Weblogic begat Fusion middleware All Based on SQL 11
  12. Application Architecture Permanent Storage Media (Disk) Memory CPU Oracle BEA

    JD Edwards, PeopleSoft, Siebel 12
  13. Evolution of Databases • 1970: Codd develops relational model, SQL

    • 1978: Oracle V1-- 128K memory • 1983: PC XT—10 MB drive • 1984: Oracle V5 – 512K memory • 2012: The same SQL? 13
  14. Tables and Columns • Structure defines relationships • Direct naming

    limited to tables, columns • Table and column names are in metadata Effect: Small namespaces • Refer to metadata to extract working data (join, select) • Working data limited in size 14
  15. We are Patching Application Structure Old Patches: • Oracle: Oracle

    Releases Oracle TimesTen In-Memory Database 11g Release 2. • SAP: HANA: The Next Wave of In-Memory Computing Technology New Patches • Memcached Memcached is an in-memory key-value store for small chunks of arbitrary data (strings, objects) from results of database calls, API calls, or page rendering. 15
  16. NoSQL offers varying approaches • Often based on key-value indices

    • Often no fixed table or column structure • Usually weaker consistency • Often optimized for append and retrieve • Exploits larger memory, distributed storage • Takes us past historic limits • But there is no uniform approach 16
  17. With JSON as the document format. Error records might look

    like this: { “ID”: 1, “ERR”: “Out of Memory”, “TIME”: “2004-09-16T23:59:58.75”, “DC”: “NYC”, “NUM”: “212-223-2332” } { “ID”: 2, “ERR”: “ECC Error”, “TIME”: “2004-09-16T23:59:59.00”, “DC”: “NYC”, “NUM”: “212-223-2332” } • Denormalized and self-contained • Requires referential ID • Requires and exploits a large name space 17
  18. Opensource Tools • Hadoop A distributed file system that provides

    high-throughput access • HBase™: A scalable, distributed database that supports structured data • Cassandra™: A scalable database with no single points of failure. • MongoDB™: A NoSQL database that optimizes writes • Hive™: A data warehouse infrastructure that provides data summarization and ad hoc querying. 18
  19. Controversies • Write-Optimization • Consistency • Permanence • What kind

    of data are you storing, and how? – Internal State: Locks and tables – External State: Account values – Big Data: Who dun what when 19
  20. Controversies • Not all access types are distinguished in the

    controversy – http://www.youtube.com/watch?v=b2F-DItXtZs 20 Note: A lot of big data concepts come from Amazon, Google, and Facebook. MySQL is a free relational database.
  21. New approaches with old assumptions – Legacy Interfaces are good

    for short-term market traction • Emulate an old interface, and your startup gets volume – But they are not optimal at a system level • Old interfaces were designed for old constraints – Will ultimately yield to completely new software architectures 21
  22. Application Architecture Permanent Storage Media (Disk) Memory CPU SQL Database

    Middleware Applications: HR, Finance, Sales, Service… 22 SQ L Block Device
  23. 23 Exploits larger RAM Assumes secondary storage is far from

    RAM
  24. Still doing key lookups in main memory 24

  25. Flash Improves Storage • Brings secondary storage closer to RAM,

    processor – IO times drop from 2 milliseconds to 20 microseconds • 2000 microseconds down to 20 – 100x improvement • There is processing power close to the device • Several “startups” fit flash to historic interfaces – XIO – Fusion IO – Pure Storage 25
  26. 26 Flash fronting for disks

  27. 27 Optimizing to legacy interfaces

  28. Application Architecture Permanent Storage Media (Disk) Memory CPU SQL Database

    Middleware Applications: HR, Finance, Sales, Service… 28
  29. A New Application Architecture Flash Storage Memory CPU No SQL

    Database New Middleware Applications: HR, Finance, Sales, Service… 29 Storage CPU
  30. NoSQL direct to Flash How much faster could real-world key-value

    stores (typical in NoSQL databases) go if their indices could be updated in non-volatile memory and not block while waiting on kernel I/O? Fusion IO’s Brent Compton, Blog, Jan 18 2012 30
  31. A New Application Architecture Flash Storage Memory CPU No SQL

    Database New Middleware Applications: HR, Finance, Sales, Service… 31 Storage CPU All New All unstable
  32. Conclusion • With MultiCore, we have plenty of CPU now

    – It’s deployed in specialized controllers – It’s available for general processing • Storage is a problem – But Flash brings it closer to the CPU • But we are limited by legacy application structure and SQL – Block device architecture limits access to logic closer to storage – NoSQL offers paths – Middleware has yet to emerge • Major players have not kept up – Oracle, SAP, Microsoft will have to adapt or acquire • The world is full of opportunity 32
  33. 33 Collaboration Applications

  34. 34

  35. 35

  36. 36 Communications

  37. 37

  38. 38

  39. 39 SaaS IT

  40. 40

  41. 41

  42. 42 Big Data Applications

  43. 43

  44. Enterprise and Infrastructure Startups • Storage – Fusion IO –

    XIO – Pure Storage • NoSQL – Couchbase – RethinkDB • Collaboration – Jive – Moxie 44 • Communications – Aeris • SaaS IT – Mobile Iron – Zenprise • SaaS VOIP – Ring Central • Big Data Apps – C3