AFDS 2012 Keynote Phillip Pokorny

Embed Size (px)

Citation preview

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    1/22

    HPC COMPUTING,FROM THE INSIDE OUT

    An Insiders Perspective Philip Pokorny, CTOPenguin Computing, Inc.

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    2/22

    2 | HPC Computing, from the Inside Out | June 14, 2012

    PENGUIN COMPUTING HOW DID WE GET HERE?

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    3/22

    3 | HPC Computing, from the Inside Out | June 14, 2012

    BEGINNING AT THE END Worlds First APU Cluster

    Delivered September 2011

    Collaboration among AMD, Sandia National Lab,Penguin Computing

    104 Altus 2A00 nodesA8-3850 APU, 16GB of memory, SSD scratchdrive, Qlogic Infiniband

    Complete solution with:Management nodes (boot, login,

    storage gateways)Networking (Ethernet and Infiniband)Six racksPower distribution with per outlet switchingSerial console monitoring

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    4/22

    4 | HPC Computing, from the Inside Out | June 14, 2012

    Founded in 1998 by Sam Ockman

    Focus on systems running Linux

    Focus on reliable systems

    Customized systems

    First Linux 1U rackmount server

    Short depth 1U

    2U Storage/Virtual Tape server with sixIDE drives

    Customer specified filesystem layout

    Customer specified RAID configurations

    HISTORY OF PENGUIN COMPUTING

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    5/225 | HPC Computing, from the Inside Out | June 14, 2012

    MEMORABLE ADVERTISING CAMPAIGNS

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    6/226 | HPC Computing, from the Inside Out | June 14, 2012

    AMD USHERS IN THE 64-BIT PC ERA

    AMD Opteron Processor AMD64 architecture64-bit registers (and more of them)

    Orthogonal instruction setIntegrated memory controllerHyperTransport Interconnect

    Linux64-bit ready and tested

    Altus 1000eDual socket systemFan control and monitoringTemperature sensor driver(written and contributed to kernel)Penguin Remote Control

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    7/227 | HPC Computing, from the Inside Out | June 14, 2012

    FROM SERVERS TO CLUSTERS

    Large orders for Altus 1000eBeowulf style clusters

    Scyld SoftwareDon BeckerMerged with Penguin Computing in 2003

    Complete clusters with all the accessoriesRacks, UPS, keyboard & monitorNetworks, cabling, management

    Factory integrated and testedSoftware installed and configuredTested as a complete clusterShipped in a crateReady to plug into network and power on-site

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    8/228 | HPC Computing, from the Inside Out | June 14, 2012

    FROM THE INSIDE OUT

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    9/229 | HPC Computing, from the Inside Out | June 14, 2012

    Dual Socket AMD Opteron Processor Memory Configurations

    CPU CORES AND MEMORY SIZES OVER TIME

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    10/22

    10 | HPC Computing, from the Inside Out | June 14, 2012

    Dual Socket AMD Opteron Processor Memory Configurations

    MEMORY PER CORE OVER TIME

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    11/22

    11 | HPC Computing, from the Inside Out | June 14, 2012

    Dual Socket AMD Opteron Processor Memory Configurations

    WHAT'S THE TREND?

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    12/22

    12 | HPC Computing, from the Inside Out | June 14, 2012

    Dual Socket AMD Opteron Processor Memory Configurations

    LONG-TERM TRENDS IN CPU AND MEMORY

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    13/22

    13 | HPC Computing, from the Inside Out | June 14, 2012

    MEMORY BANDWIDTH OVER TIME

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    14/22

    14 | HPC Computing, from the Inside Out | June 14, 2012

    THOUGHTS ON THE FUTURE

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    15/22

    15 | HPC Computing, from the Inside Out | June 14, 2012

    STORAGE PERFORMANCE

    7200 RPM drives from 2002 to 2012

    80 GB to 3000 GB

    40 MB/sec to 160 MB/sec

    Lots more space, not much more performance

    RAID

    Less than perfect scaling

    Reliability level impacts performance

    Motherboard SATA ports limit number of drives

    Solid State Drives

    Limited writes

    May need ATA TRIM for better results

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    16/22

    16 | HPC Computing, from the Inside Out | June 14, 2012

    STORAGE PERFORMANCE

    SATA really? Compatible with 1985 MFM drive controller!

    Assumes compatibility at register level

    Butquirks to every controller One drive per physical portHot swap added onOops, forgot the drive activity signal

    SASHigher performance drives (15,000 RPM)Flexible number of drivers per controllerCan host SATA drivesMultiple ports for redundancy and performance

    Ask for it

    IS THAT A GPU IN YOUR SOCKET?

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    17/22

    17 | HPC Computing, from the Inside Out | June 14, 2012

    IS THAT A GPU IN YOUR SOCKET?

    Mixed number of CPU/GPU

    Socket compatible FPGAs Spansion flash controller and NOR flash DIMMs

    Just need the hardware

    POWER EFFICIENCY AND IDEAS FOR IMPROVEMENT

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    18/22

    18 | HPC Computing, from the Inside Out | June 14, 2012

    POWER EFFICIENCY AND IDEAS FOR IMPROVEMENT

    CPUs well-covered elsewhere

    MotherboardsMore efficient voltage converters (VRM)Better VRM tuningGenerate voltages only where needed

    Power supplyHigher efficiency

    Beware of hidden power distribution boards

    Single rail (one voltage, not five or more)

    Common infrastructure12.5V distributionShared fansMatch power supplies to server load

    WHERE DOES THE POWER GO?

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    19/22

    19 | HPC Computing, from the Inside Out | June 14, 2012

    WHERE DOES THE POWER GO?

    You cant control what you cant measure

    Need to compare multiple systems

    Need high-speed capture for some analysis

    Diagnose problems

    Spot trends

    CPU versus motherboard versus anything else

    Monitor each voltage rail individuallyCPU, motherboard, fans, hard drives

    High speed (1000 samples per second or faster)

    Correlate samples to code execution

    COMING SOON

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    20/22

    20 | HPC Computing, from the Inside Out | June 14, 2012

    COMING SOON

    A glimpse of what we might learn

    CPU power draw during make j4

    Different phases of compilation

    Clearly different power levels

    Larger files can load CPU 100% forlong periods of time

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    21/22

    21 | HPC Computing, from the Inside Out | June 14, 2012

    THANK YOU QUESTIONS?

  • 7/31/2019 AFDS 2012 Keynote Phillip Pokorny

    22/22