TECHNION MEETING on CLUSTER COMPUTING

1
TECHNION MEETINGon CLUSTER COMPUTING

• Anne Weill-Zrahia
• Technion Computer Center

2
Programme

• Introduction
• Round table of participants
• Future subjects and activities

3
Introduction

• Cluster an ensemble of independently
  operational elements integrated in some medium
  for cooperative and coordinated behaviour
• Computer cluster processing components
  integrated by an interconnection network and
  cooperating on a common application program or
  workload
• Beowulf cluster computer cluster made of PCs,
  managed by low-cost software so as to maximize
  the prize-performance.

4
Building blocks for a cluster

• Compute nodes 1 or more processors, RAM
  memory, hard disk
• Network NICs, switches,channels
• Libraries for message-passing applications
• Software environment for management of resources

5
Cluster Classification

• Dedicated/ non dedicated nodes
• HPC (High Performance Computing)vs HA ( high
  availability)
• Node hardware PC, workstation
• Node system Linux, Windows NT
• Homogeneous/heterogeneous clusters

6
HPC Computing
7
HPC Today

• 47 systems in Top500 are more than 1 Tflop.
• First place Earth Simulator , 35 Tflops
• 2 PC based clusters in the 10 top spots
• 55 Intel based and 8 AMD-based clusters in the
  TOP 500
• 14 are self-made

8
Cluster Computing now

• Processors available are more powerful (gt1 Ghz)
• Advances in networks
• Price/performance is still good
• Low-cost upgrade
• Future grid computing

9
Disadvantages of clusters

• Programming is more difficult
• Latency of networks is still higher than in
  supercomputers
• Lack of management and performance evaluation
  tools

10
Programming models

• Sharing cluster resources while running serial
  codes
• Parallel programming - at node level(OPEN-MP)
• Parallel programming across nodes (MPI)
• Both approaches

11
MPI Programming

• Different versions of MPI
• How to ensure speedup
• How to get load balancing
• Performance monitoring
• Debugging

12
Cluster Management

• High-performance CPU cycles/second under ideal
  circumstances
• High-throughput CPU cycles/day under non-ideal
  circumstances

13
Conclusions

• Clusters are a good solution for HPC in terms of
  price/performance
• They are not yet easy to use as traditional
  supercomputers
• They can open the way to grid computing

14
Cluster computing at TCC

• Clinux -
• 6 Pentium III , 866 Mhz, 1 Gb RAM,CC Ethernet
  switch
• Kelp -
• 8 dual Pentium III , 800 Mhz, 512 Mb RAM, Fast
  Ethernet switch

15
Possible topics

• MPI programming
• Cluster management
• Benchmarking
• Developing code on clusters