Please enable JavaScript.

Coggle requires JavaScript to display documents.

Week 4 Distributed (Clusters (Performance (Application (Runtime comparison…

- - - - New tech is to use PCI Express
      - Gigabit ethernet used but not that much faster when using TCP/IP. Because of processing delays. These are queueing systems
    - - cLAN by VIA company
        
        Very low latency
      - QsNet
      - ANSI/VITA
      - Infiniband - Widely used
  - - - Latency measured by timing
      - Bandwidth by benchmarking
      - Network conditions change so hard to rely on
      - Done on lightly loaded machines for best achievable performance
    - - Maximum aggregate performance
        P = NCFR
        
        Can use FLOPS or MOPS
    - - Calculate MOPS or MFLOPS divided by runtime
      - Computed by running a similar application to the intended one
      - More meaningful than theoretical peak performance
      - Need to estimate No. floating point ops in the code
      - Optimize algorithm
      - Problem when code was tuned for a particular platform
      - Runtime comparison
        
        Running the same application in several different systems and comparing the wall-clock time it takes
        
        Performance tuning could distort the results
        
        Good thing is no need for benchmarking code
    - - T(1)/T(N)
        Where T(N) is the wall clock time of N processors
      - Close to N means scales well
      - Helps us decide what the best no. processors is
      - Inefficient single CPU version can falsify results
        
        Memory bottleneck
    - - E = P(N)/N
        Where P(N) is performance on N processors (probs amdahl)
      - Same problems as scalability
    - - Shows the extent and app is making use of the available performance
      - Depends on the type of app
    - - System level effects
        
        Competition between tasks
        
        I/O Contention
        
        Memory swapping
        
        Job scheduler inefficiency
        
        Job startup delay
      - Long term effects mesaured by assessing effects
      - Tools for assessment
        
        sar
        
        vmstat
        
        netstat
        
        iostat
    - - Widely used
      - Test aggregated bandwidth and latency of IPC network
      - API written in C and MPI installed
    - - High performance LINPACK (HPL) benchmark is widely used for clusters
      - Can overestimate performance
      - Library of functions for solving linear equations and linear least-squares problems
      - FORTRAN
      - Single or multi-node clusters
    - - 5 general computational kernels
      - 3 simulated fluid dynamics
      - Can be used for single or multi-node clusters
      - MPI recommended
- - - - Job startup
      - Security