Supercomputing over a network

1
Supercomputing over a network

• Abhinav Jalan
• Retesh Chadha

2
Goals accomplished

• Most cost effective supercomputer.
• Utilization of the unused computing power on the
  network to the maximum extent.
• Most easy to program supercomputer.

3
A framework that

• can run in parallel a sequential program.
• automatically handles all the intricacies of
  synchronization and communication.
• automatically recovers from any failure in the
  system.
• ensures automatic load balancing in the network.
• automatically detects and uses any resource on a
  network.
• easy to program.

4
Running in parallel a sequential program

• Automatically fragments a sequential program into
  a parallel one to the maximum extent possible.
• Runs any procedure of a sequential program in
  parallel to main program.
• No parallel algorithm needed.

5
Synchronization and communication

• Implicitly handles synchronization among parallel
  running fragments.
• All communication among process is implicit.
• All inter machine communication is non blocking
  and done in parallel.
• Ensures the integrity of data.

6
Automatic Failure Recovery

• Automatically detects a failure.
• Automatically restores the system back.
• Restores the system by transferring the jobs from
  crashed system to others.

7
Fault tolerance
Job 1
Job 1
Job 2
Sending Job
Returning Job
8
Fault tolerance cont
Job 2
Sending Job
Returning Job
9
Fault tolerance cont
Job 2
Job 2
Sending Job
Returning Job
10
Automatic load balancing

• Job is distributed to the best available machine
  on the network.
• Best available machine is chosen according to the
  load of the machine where load comprises of
• CPU usage
• Number of threads in ready state
• CPU clock speed
• Memory available
• Number of processors
• Ensures no node is overburdened.

11
Automatic detection of resource

• Dynamically extendible.
• Makes it easy to deploy anywhere.
• Detects a new node as and when added to the
  network and start utilizing it.

12
Easy to program

• No need to write parallel code.
• Easy to learn
• No extra libraries needed to be learned
• No need to write code to handle fault tolerance,
  its automatic.
• No need to specify nodes to be used.

13
Features of class library

• Easy to learn and use
• Platform independent
• Object oriented
• Robust
• Extensible
• Rich class library

14
Other features cont

• Handles changes in object linkage structure.
• All communication is non blocking in separate
  threads.
• IPC is highly fast as it takes places in
  parallel.
• Threads are reused.

15
Competitive Analysis

• Super Computers
• Cray, IBM, CDAC
• Similar libraries available
• MPI (Message Passing Interface)
• PVM (Parallel Virtual Machine)
• Distributed Operating Systems
• Not many implementations available

16
Development process

• Development of a domain-neutral class library.
• Design of a framework for distributed computing
  using it.
• Redesign of the framework to accommodate fault
  tolerance.
• Development of an application for
• automatic resource detection
• load balancing
• fault detection.

17
The project

• This is the only system of its kind.
• Has no dependencies, has been done from scratch.
• The idea was coined by us.
• 20,000 lines of code.

18
The project and IIIT

• IIIT has the potential to house Indias fastest
  supercomputer, faster than PARAM10000.
• IIIT can offer super computing service to other
  institutes.

19
What more can be added?

• Extension to Internet
• Security
• Extension to other platforms
• Support for standard algorithms and
  decompositions
• Output redirection

20
Limitations of the Model

• Not very fast for multiprocessor computers
• All machines must be on same platform
• High resource requirement
• Synchronous objects and jobs take large memory
• Uses too many threads

21
Thank you
FOR MORE INFO...
Project report available in library and at
http//www.iiita.ac.in/IIITSupercomputer/report.
doc