A Technology Analysis of High Performance Computing

1
A Technology Analysis ofHigh Performance
Computing
hpc_at_etsu
2
Agenda

• introducing HPC
• the technology market
• our involvement
• a 4-layer architecture
• project methodology
• results from case studies
• cost considerations
• future opportunities
• concluding remarks

3
HPC Technology Overview

• superior computing through advanced hardware and
  software configurations
• traditionally the domain of large scale (SMP,
  MPP) supercomputers
• giving way to clusters of smaller machines
• two approaches to clusters
• dedicated
• shared-use (i.e., cycle-scavenging)
• not limited to just research projects

4
HPC in the Marketplace
5
Benefits of HPC Computing

• speed
• effectiveness of HPC clusters often measured in
  terms of speedup
• accuracy
• when time is limited, researchers often use
  faster, but less accurate, algorithms HPC
  performance increases are often used to reclaim
  lost accuracy
• timeframe
• researchers often decide the scope of their
  research according to what they can accomplish in
  a given timeframe

6
Our Involvement

• CSCI capstone projects
• 3-semester, strategic learning experience
• small group project as alternate to thesis
• HPC capstone project
• an exploratory project to
• investigate and report on promising HPC
  technologies
• identify and assist ETSU clients to leverage HPC
• track labor costs
• transition capstone to another team
• 9 graduate students

7
Mission Objectives

• evaluate lower-cost, contemporary strategies for
  implementing HPC at ETSU
• investigate the feasibility and cost
  effectiveness of the different strategies
• determine impediments to deployment within the
  ETSU infrastructure
• transfer expertise to a 2nd generation of
  capstone students

8
The 4 Layers of HPC Environments

• Application
• Middleware
• Infrastructure
• Hardware

9
Hardware Layer

• computing machines
• number of processors
• type of processors (32-bit vs. 64-bit)
• memory
• networking infrastructure
• topology
• throughput
• latency
• storage devices
• shared file systems
• local file systems
• global file access requirements

10
Hardware Layer

• 20-node benchmark cluster in WW-006
• dedicated cluster
• combination of remote and local file systems
• remote boot
• local workspace
• Linux
• 10-node alpha cluster in WW-002
• shared-use cluster
• Windows
• 100 Mb/s Full Duplex, VLAN configuration

11
Infrastructure Layer

• system software
• operating system
• disk-based (Windows, Linux)
• diskless (Linux)
• software tools (compilers, libraries)
• infrastructure services
• network services (AD, DNS, DHCP)
• remote access protocols (SSH)
• imaging tools (Ghost, DRBL, RIS, Kickstart)

12
Middleware Layer

• extend the software system to create the illusion
  of a single, unified computing environment
• batch schedulers / resource managers
• assign processes to host systems
• manage access to shared resources
• virtual computing machine
• facilitate inter-process communications
• system tools
• provide console-like, distributed command
  execution
• extends reporting capabilities

13
Middleware Layer

• Condor (shared-use strategy)
• a batch scheduler and a resource manager
• used for dedicated clusters or cycle-scavenging
• multi-platform UNIX or Windows
• MPICH2 (dedicated strategy)
• an implementation of the Message Passing
  Interface (MPI)
• inter-processor communication
• multi-platform UNIX or Windows

14
Application Layer

• applications that transfer readily to HPC
  environments
• embarrassingly parallel applications
• break application into independent pieces
• process each piece independently
• assemble results
• explicitly parallel applications
• structure application as set of cooperating
  programs
• minimize overhead to maximize work
• end-user applications
• rendering animations
• microlensing
• galaxy collisions

15
Project Methodology

• Literature Research
• Technology Evaluations
• Case Studies

16
Literature Research

• HPC is new, and very much fluid
• dynamic complex environments
• evolving technologies configurations
• no such thing as a novice HPC user
• only expert or advanced
• technical documentation
• not totally accurate, not as advertised, not
  advertised at all, cant read advertisement
• little in way of business case reasoning

17
Technology Evaluations

• hardware
• test clusters
• alpha test lab
• software infrastructure
• system imaging vs. remote booting
• environment emulation
• middleware
• Condor
• MPICH2
• applications
• roll your own

18
Case Studies

• digital media animation rendering
• shared-use environment
• Windows
• physics microlensing
• dedicated computing environment
• physics galaxy collision modeling
• dedicated computing environment
• Linux/Windows interoperability

19
Results from Case Studieswith ETSU Clients

• Maya Animation Rendering
• Microlensing
• Galaxy Collision Modeling

20
Maya Animation RenderingDr. Cher Cornett, Dr.
Marty Fitzgerald Digital Media

• standard animation rendering
• serial rendering process
• synchronous, manual effort
• processor storage intensive
• parallel animation rendering
• distributed rendering process
• batch, automatic procedure
• cycle-scavenging

21
Results

• time comparison of Maya serial vs. Maya in
  parallel near-linear speedup with Condor
• Baked Explosion Sink rendering projects
• recent alpha testing

22
Microlensing Dr. Richard Ignace Physics
Department

• serial application (MC Lens)
• Monte Carlo simulation for gravitational lensing
• parallelization of MC Lens
• incremental approach to reengineering
• results of the parallelized version are pending

23
Galaxy Collision Modeling Dr. Beverly Smith
Physics Department

• automatic galaxy collision (AGC) simulation
• 3-body interaction model
• serial genetic algorithm (Pikaia)
• processor storage intensive
• parallel automatic galaxy collision (PAGC)
  program
• parallel genetic algorithm (Parallel Pikaia)
• distributed random number generator

24
Results

• parallelization of AGC yielded PAGC
• revision of Parallel Pikaia yielded PPGA
• near-linear (12.8x) speedup with PAGC

25
CostConsiderations

• Supporting HPC
• HPC Hidden Costs
• Timesheet Analysis

26
Supporting HPC

• training
• four weeks to train HPC2 team
• eight weeks for HPC2 to become proficient
• developing and maintaining infrastructure
• 100 hours over two semesters
• reverse engineering
• 750 hours for PAGC (AGC 4000 SLOC)
• 250 hours for MC Lens ( 800 SLOC)
• maintenance of converted applications
• developing and maintaining expertise
• two to three times longer than expected

27
HPC Hidden Costs

• software technologies
• free vs. commercial software
• system architecture
• dedicated vs. shared-use
• increased power consumption
• hardware software maintenance costs
• system administration overhead
• steep learning-curve for programmers and
  scientists alike
• security concerns
• clusters are notoriously difficult to secure

28
Timesheet Analysis
29
FutureOpportunities

• Other ETSU Clients
• Additional Package Evaluations
• Software Development

30
Other Potential ETSU Clients

• researchers with HPC
• Dr. Scott Kirkby, Chemistry
• Dr. Jeff Knisley, Mathematics
• researchers indicating interest
• Dr. Debra Knisley, Quantitative Biology
• Dr. Cecilia McIntosh, Biochemistry
• Quillen College of Medicine

31
Additional Package Evaluations

• Moab
• job scheduler and policy engine
• graphical management and monitoring
• web-based job submission
• RIS with partitioned source image
• Kickstart for configured image installation
• Symantec Ghost for automated imaging

32
Software Upgrades

• PPGA, support for
• additional parameters
• additional strategies for GA-based optimization
• multi-threading to improve algorithm performance
• HPC_Render
• GUI interface
• dynamic division of Maya files

33
Hardware Upgrades

• 20-node cluster in WW-006 to 60 nodes
• additional network components
• switch ports
• patch panels
• network cables
• internal bandwidth
• 10/100 Mb to GigE or Infiniband
• explore other hardware configurations
• storage-node architecture
• dedicated network paths

34
Conclusions

• Accomplishments
• QA

35
Accomplishments

• implemented and documented two strategies for
  achieving HPC at ETSU
• three case studies demonstrating benefits
• developed A Technology Analysis of High
  Performance Computing
• feasibility, challenges, labor cost of deploying
  HPC in campus network
• technology overviews and how-to guides for
  open-source packages, infrastructure issues
• helped to bring HPC to ETSU
• enhance ETSUs standing in academic research
• increase ROI for underutilized computing assets

36
QA

• Visit hpc_at_etsu Online
• http//cscidbw.etsu.edu/hpc