High Performance Computing

1
High Performance Computing Supercomputers

• Robert Whitten Jr

2
Welcome!

• Todays Agenda
• Questions from last week
• Parallelism
• Supercomputers

3
What is a supercomputer?

• The fastest computer at that moment
• Top 500 list is updated every 6 months
• Serial vs. parallel
• Serial means youre doing one thing at a time
  sequentially
• Parallel means youre doing multiple things at
  the same time
• Parallel is the way to do things

4
Parallelism

• Performing multiple tasks simultaneously will
  increase how much work can be done in a given
  amount of time

5
Parallel example

• How long would it take a person to build a car?
• Assume they are a pretty good mechanic
• Assume they could build a car in a month by
  themselves
• Now were talking about serial processing
• What if 2 mechanics worked on the car?
• What if 3?...4?
• Now were talking parallel processing

6
Shared Memory Parallelism

• If mechanics used the same bucket of bolts
• Contention occurs whenever someone reached for
  the same bolt as someone else
• Communication occurs when bolts (data) need to be
  passed back and forth between mechanics

7
Distributed memory parallelism

• Now each mechanic has his own section of the car
  to work on
• Problem decomposition
• Sharing of a common resource is not a major
  factor now
• What if one section is easier to do?
• What if one mechanic is faster than the others?
• Load balancing occurs when work has to be
  distributed among other mechanics

8
Supercomputers

• Distributed systems
• Teragrid
• SETI_at_home
• Clusters
• Each nodes is an individual computer
• Link them all together and youve got a cluster
• Supercomputers
• Proprietary (Cray, IBM, etc)
• Custom interconnect networks

9
Distributed systems
10
Distributed systems

• Typically heterogeneous systems
• PCs, Macs, Linux, etc
• All nodes are physically separated from other
  nodes
• Geographically
• Logically
• Typically only communicate data back and forth
• Works best if data can be divided into
  independent chucks with little interdependence

11
Clusters
12
Clusters

• Typically homogeneous
• Might be some difference in hardware, but minimal
  
• Typically co-located
• Share common network
• Ethernet, infiniband, etc

13
Supercomputers
14
Supercomputers

• Typically homogenous
• Exceptions are out there (i.e. Roadrunner _at_ LANL)
• Share common network fabric
• Interconnect between processors
• Interconnect between nodes
• Nodes cannot be independent of each other
• Service nodes
• Login nodes
• I/O nodes
• Compute nodes

15
Jobs

• Execution object on a distributed system
• Can be interactive or batch
• Interactive means a user has to be present to
  enter data
• Batch means data is read from files and user does
  not have to be present
• Allows for greater utilization of the machine
  since many jobs can be submitted at same time

16
Homework

• Send me that email if you havent already
• whittenrm1_at_ornl.gov

17
Questions?
http//www.nccs.gov
Oak Ridge National Laboratory
U. S. Department Of Energy 17