Presented by: Prof Mark Baker

1
Emerging Technologies and Ideas 

• Presented by Prof Mark Baker
• ACET, University of Reading Tel [phone]
  8615 E-mail Mark.Baker_at_computer.org
• Web http//acet.rdg.ac.uk/mab

2
Overview

• Aims and Objectives of Workshop.
• Emerging Technologies
• Multi-core systems,
• Cloud-based systems,
• Virtualisation,
• Green computing.
• Embracing failure.

3
Thanks

• Would like to say thanks for everyone attending
  and thanks for Guy Robinson for helping me
  organse this event.
• Attendees
• ECMWF,
• Met Office,
• Met Dept, University of Reading,
• ACET, University of Reading,
• NAG,
• Cray,
• IBM,
• Leeds,
• Oxford,
• Daresbury, STFC.

4
Aims and Objectives of Workshop.

• Meteorological community have a very long history
  of developing and using large and complex climate
  and weather modelling codes.
• The codes used have been developed over many
  years and have required hundreds of years of
  person effort to achieve the scientific quality
  that can be obtained today.
• The Fortran programs are based on legacy codes,
  and are sensitive to compilers, libraries, and
  the execution environment.
• Code formulation issues can create inertia to
  exploiting new computing technologies.

5
Aims and Objectives of Workshop.

• We wish explore the gap between the emerging
  technologies and innovations as well as the way
  these can possibly be utilised by the
  meteorological community.
• In particular we want to discuss and debate the
  codes that are used to model climate/weather, and
  the emerging technologies and innovations
• Other issues will emerge from the discussion of
  new emerging technologies that may be relevant to
  the meteorological community such as robustness
  and reliability too.

6
Schedule

• 1030 1100 - Introduction and Overview of
  Emerging Technologies and Ideas -MAB
• 1100 1120 - IFS program developments for future
  systems ECMWF
• 1120 1150 - Using earth system models on UK
  academic computers a study of generic
  performance issues Met Dept.
• 1150 1200 - Exploring earth system model
  performance to anticipate change in emerging
  computer architectures Met Dept.
• 1200 1210 - METAFOR, Climate Metadata for
  Climate Modelling Digital Repositories Met
  Dept.
• 1210 1230 - If only we could forecast computer
  architectures as well as we can forecast the
  weather! Met Office
• 1230 1300 - Bridging the Gap. We are on the
  brink of a new era if only...! IBM
• 1300 1400 Lunch
• 1400 1430 - What can HECToR do for me? NAG
• 1430 1500 - Computational and Numerical
  Challenges in Climate and Environmental Modelling
  ACET
• 1500 1510 - Advanced Scalable Algorithms for
  Advanced Architectures ACET
• 1510 1525 - The HPC-NA Roadmap Activity -
  Looking into the Future of Numerical Advanced
  Computing Oxford
• 1525 1600 - Panel Discussion Session

7
Multi-core Systems

• The number, scale and type multi-core systems
  available are changing the landscape of
  application development.
• Parallel computing used to be the preserve of a
  relatively small group of highly skilled
  programmers
• The emergence of multi-core systems means that
  parallel programming is becoming a mainstream
  activity again!
• It is important to ensure that multi-core
  machines can be effectively and efficiently
  programmed.
• Also, there is an issue of how to program large
  clusters of multi-core processes where there is
  the need for intra and inter communications.
• A key part of that strategy is the availability
  of high-level tools and libraries to assist
  expert and the novice parallel programmers create
  successful programs.

8
Some Issues

• Can expect 10s-100s of core on processors in the
  coming years
• Currently vendors seem to believe that
  multi-threaded programs will be OK for
  programming these systems
• Also, vendors seem to think a shared memory
  programming model is fine!
• For some programs this may be OK, but coming from
  a HPC/parallel computing arena there are issues
  that threads do not address.
• Architecture of multi-core processors are
  different use different cache levels to share
  data
• Consequently will need different strategies for
  using different processor types.
• Need to create more efficient applications
  rather than 20 efficiency, want 90 (part of
  the green computing revolution).

9
Thoughts

• On large systems you will want to ensure data
  locality so you need to have thread affinity.
• Want to effectively load-balance applications
  running on these systems.
• Need to go back and seriously think about
  concurrent programming course and fine grain
  approaches!
• May want to synchronise threads, (all or groups).
• Will want to optimise strategies for using cache,
  this may depend on 2nd or 3rd level cache!
• Want to overlap communication and computation
  too.
• Need to consider the limitations of the
  architecture when considering communications and
  access to the system bus.

10
Thoughts

• Need to cope with different operating systems,
  ranging from UNIX to Microsoft Windows
• May need to work with different programming
  paradigms message passing (e.g. MPI), OpenMP,
  through to data - parallelism (e.g. UPC and
  Fortran) which best suit the applications
  needs!!
• In the future will need to consider heterogeneous
  multi-core processors.
• Think about mixed precision programming single
  and double precision.
• Lighter-weight threads
• Need to schedule program efficiently on large
  machines do not want fragmented partitioning

11
Multi-Core Some Offerings

• AMD
• Opteron
• Athlon 64
• Turion 64
• Barcelona
• ARM
• MPCore (ARM9 and ARM11)
• Broadcom
• SiByte
• Cradle Technologies
• DSP processor
• Cavium Networks
• Octeon (16 MIPS cores)
• IBM
• Cell (Sony, and Toshiba)
• POWER4,5,6
• Intel
• Core
• Xeon

• Motorola
• Freescale dual-core PowerPC
• Picochip
• DSP devices (300 16-bit processor MIMD cores on
  one die)
• Parallax - Propeller (eight 32 bit cores)
• HP - PA-RISC
• Raza Microelectronics - XLR (eight MIPS cores)
• Stream Processors
• Storm-1 (2 MIPS CPUs and DSP)
• Sun Microsystems
• UltraSPARC IV,
• IntellaSys - seaForth-24.

12
AMD and Intel Multi-core Processors
Quad-Core Clovertown
Core 1
Core 2
Core 3
Core 4
100011
L2
L2
Front-Side Bus
Front-Side Bus
Memory Controller
Northbridge

• AMD Independent L2 caches
• Multiple Memory modules,
• Communication over Point-to-
• Point HyperTransport Channels.

• Intel Shared L2 caches
• Single Memory,
• Communication over Front- side Buses.

13
Cell/B.E. Architecture

• Power Processing Element (PPE)
• Eight Processing Elements (SPE)
• 4 SIMD ALUs
• DMA Engines
• 256 kB Local Storage (LS)
• System Memory
• 25 GB/s
• Element Interconnect Bus (EIB)
• Over 200 GB/s

14
GPUs - GeForce 8800

• 16 threaded SMs, gt128 FPUs, 367 GFLOPS, 768 MB
  DRAM, 86.4 GB/S Mem BW, 4GB/S BW to CPU

15
Multi-core Issues to Solve

• Communications, synchronsation, resource
  management between/among cores.
• Debugging connectivity and synchronisation.
• Distributed power management.
• Concurrent programming.
• OS virtualisation.
• Modelling and simulation.
• Load balancing.
• Algorithm partitioning.
• Performance analysis.

16
Reason for the crisis

• Memory bandwidth is a major limitation with
  multi-core systems
• All cores must share access to the same memory
  and therefore have to take turns to read from or
  write it.
• Disk speed can throttle performance
• If a disk can read only 40 MBytes/sec, then a
  program that needs to read a 400 MByte file will
  take a minimum of 10 seconds to open it, even if
  it has no computation to do. Software and
  scalability
• Software today is frequently not written to take
  advantage of multiple cores efficiently
• It is technically challenging to
  write efficient code that is free of bugs a
  highly scalable program is difficult to build and
  requires a lot of expertise.

17
There is MPI style messaging and ..

• OpenMP annotation or Automatic Parallelism of
  existing software is practical way to use those
  cores with existing code
• As parallelism is typically not expressed
  precisely, one needs skill to get good
  performance
• Writing in Fortran, C, C, Java throws away
  information about parallelism
• HPCS Languages should be able to properly express
  parallelism but we do not know how efficient and
  reliable compilers will be
• High Performance Fortran failed as language
  expressed a subset of parallelism and compilers
  did not give predictable performance.

18
There is MPI style messaging and ..

• PGAS (Partitioned Global Address Space) like UPC,
  Co-array Fortran, Titanium, HPJava
• One decomposes application into parts and writes
  the code for each component but use some form of
  global index.
• Compiler generates synchronisation and messaging.
• PGAS approach should work but has never been
  widely used presumably because compilers not
  mature.

19
What are Clouds?

• Clouds are Virtual Clusters
• They may cross administrative domains or may
  just be a single cluster the user cannot and
  does not want to know!
• Clouds support access (lease of) computer
  instances
• Instances accept data and job descriptions (code)
  and return results that are data and status
  flags.
• When does Cloud concept work
• Parameter searches, LHC style data analysis, Face
  Book

20
What makes a Cloud?

• Virtual Machines.
• VM Manager
• Scalability.
• File system Infrastructure.
• Remote access (portal).
• Cost?
• Fairly low cost for CPU/data movement.
• Security?

Key Parts of Cloud Definition
21
Cloud Architecture
Source S.Tai
22
Some Commercial Cloud Offerings

• Problem Commercial offerings are proprietary and
  usually not open for cloud systems research and
  development

23
Cloud Computing Service Layers
Description
Services
Services Complete business services such as
PayPal, OpenID, OAuth, Google Maps, Alexa
Application Focused
Application Cloud based software that
eliminates the need for local installation such
as Google Apps, Microsoft Online
Development Software development platforms used
to build custom cloud based applications (PAAS
SAAS) such as SalesForce
Platform Cloud based platforms, typically
provided using virtualization, such as Amazon
ECC, Sun Grid
Storage Data storage or cloud based NAS such
as CTERA, iDisk, CloudNAS
Infrastructure Focused
Hosting Physical data centers such as those run
by IBM, HP, NaviSite, etc.
24
Technical Issues about Clouds

• Using VMs.
• No common standards APIs often different on each
  system.
• Protocols Web Services and RESTful.
• Proprietary workflow systems.
• Database systems.
• Not clear if MPI/OpenMP can be run!
• Multiple VMs on a processor.
• Code development, debugging and optimisation!
• Security.
• Pricing/costs

25
Virtualization in General

• Virtual machines run in software that emulates
  computer hardware
• Host machine hardware running the virtual
  machine software,
• Host operating system operating system running
  the virtual machine software,
• Hypervisor slimmed down host operating system
  that virtualises the physical hardware,
• Guest system operating system.
• Examples of Virtual Machines
• VMware,
• Microsoft Virtual PC and Microsoft Virtual
  Server,
• Parallels Workstation,
• Sun xVM,
• Kernel-based Virtual Machine (KVM),
• Xen (Opensource),

26
Virtual Machines

• VM technology allows multiple virtual machines to
  run on a single physical machine.

App
App
App
App
App
Xen
Guest OS (Linux)
Guest OS (NetBSD)
Guest OS (Windows)
VMWare
UML
Virtual Machine Monitor (VMM) / Hypervisor
Denali
Hardware
etc.
Performance Paravirtualization (e.g. Xen) is
very close to raw physical performance
27
Virtual Machines

• Multiple virtual machine instances on a single
  physical host
• Fault tolerance,
• Isolated OS instances,
• Virtual servers.
• Use emulation to support different instruction
  set architectures such as Intel IA-32 and
  PowerPC.
• Support novel architectures.
• Support for high-level language virtual machines
  (Java).

28
Virtualization in General

• Advantages of virtual machines
• Run operating systems where the physical hardware
  is unavailable
• Easier to create new machines, backup machines,
  etc.
• Software testing using clean installs of
  operating systems and software
• Emulate more machines than are physically
  available
• Timeshare lightly loaded systems on one host
• Debug problems (suspend and resume the problem
  machine)
• Easy migration of virtual machines (shutdown
  needed or not)

29
Workload Consolidation pros/cons

• Pros
• Each application can run in a separate
  environment delivering true isolation,
• Cost Savings Power, space, cooling, hardware,
  software and management,
• Ability to run legacy applications in legacy OSs,
• Ability to run through emulation legacy
  applications in legacy HW.
• Cons
• Disk and memory footprint increase due to
  multiples OSs,
• Performance penalty caused by resource sharing
  management.
• Workload consolidation provides the basis most
  usages/benefits of virtualization

30
VMS

• VM Use
• Workload isolation,
• Workload migration,
• Workload migration for dynamic load balancing,
• Workload migration for disaster recovery,
• Control of VM resources,
• Legacy OSs
• Green computing,

31
Green Computing

• Why?
• Computer energy is often wasteful
• Leaving the computer on when not in use (CPU and
  fan consume power, screen savers consume power).
• Pollution
• Manufacturing techniques,
• Packaging,
• Disposal of computers and components.
• Toxicity
• Toxic chemicals used in the manufacturing of
  computers and components which can enter the food
  chain and water!

32
Energy Use of PCs

• CPU uses 120 Watts.
• CRT uses 150 Watts.
• 8 hours of usage, 5 days a week 562 Kwatts
• If the computer is left on all the time without
  proper power saver modes, this can lead to 1,600
  Kwatts.
• For a large institution, say a university of
  40,000 students and faculty, the power bill for
  just computers can come to 1.5 million / year
• Energy use comes from
• Electrical current to run the CPU, motherboard,
  and memory,
• Running the fan and spinning the disk(s),
• Monitor (CRTs consume more power than any other
  computer component).

33
Reducing Energy Consumption

• Turn off the computer when not in use, even if
  just for an hour.
• Turn off the monitor when not in use (as opposed
  to running a screen saver).
• Use power saver mode
• In power saver mode, the top item is not
  necessary, but screen savers use as much
  electricity as any normal processing, and the
  screen saver is not necessary on a flat panel
  display.
• Use hardware/software with the Energy Star label
• Energy Star is a seal of approval by the Energy
  Star organization of the government (the EPA)
• Use LCDs instead of CRTs as they are more power
  efficient.

34
Embracing Failure

• As more systems encompass ever-increasing numbers
  of components, even a small fault rate on
  individual processors will generate multiple
  faults across the components, stopping
  long-running applications in their tracks.
• Weather/Climate Modelling
• Outlive Complexity
• Increasingly sophisticated models,
• Model coupling,
• Inter-disciplinary.
• Sustained Performance
• Increasingly complex algorithms,
• Increasingly diverse architectures,
• Increasingly demanding applications.

35
Embracing Failure

• 1000s of core,
• 2 Gbytes Memory per core,
• Caching problems,
• File systems,
• Networks,
• Operating Systems,
• Middleware (MPI/OpenMP),
• Applications algorithms.

36
Key Challenges

• How do you know if there were any transient or
  permanent failures of the hardware or system
  software that invalidate the computation?
  (Detect)
• Can a middleware library implement the
  functionality of the computational workflow model
  isolating the computational software from the
  transient and permanent failure management?
  (Isolate/Contain)
• In the case of permanent failures, how does one
  remap the computation to the remaining available
  resources? Does the application programmer have
  to do this? (Recover)

37
Summary

• Multi-core systems need to understand
  underlying hardware, and be able to efficiently
  program clusters of multi-core processors.
• Cloud-based systems becoming increasing popular
  Grid is dead!.
• Flexible utility-based platform, that has a range
  of interesting and innovative properties.
• Virtualisation being using in Clouds, but also
  to provide management and control of machines.
• Green computing becoming important, and a key
  feature of FP7.
• Embracing failure machines are becoming bigger,
  more sophisticated and so failure is increasingly
  like must check out ways of creating reliable
  applications!