Section 1.5, The Power Wall, 1.6 Multiprocessors

• Look at Figure 1.15, page 39
• The authors tell us that we have reached a practical level for cooling processors.
• With current technology, power is consumed to perform a switch.
  • Power = Capacitive Load × Voltage² × Frequency switched.
  • Frequency Switched is directly related to clock speed.
    • Thus higher clock rates require more power in the same chip
  • Capacitive load is related to the types of transistors and wires, as well as how many transistors are connected to an output.
• Transistor leakage
  • Current that leaks through the transistor, even though the transistor is in a "off" position.
  • This becomes worse as we shrink the size of transistors to make them switch faster.
  • It also becomes worse as voltage is lowered.
  • The book says that 40% of power consumption is due to leakage.
• These problems have combined to make "commodity processors" too hard to cool as clock speeds increase.
• Options to deal with heat dissipation
  • Better and better heat sinks.
  • Turn off unused sections of the chip
  • Special purpose chips (no FP for example)
  • Multicore.
• To combat this problem, but to still sell chips, manufactures have gone to a multicore solution.
• Look at figure 1.16, page 42.
• Multicore means multiple processor per chip
• "The official plan of record for many companies is to double the number of cores per microprocessor per semiconductor technology generation, which is about every two years."
• The problem with this switch
  • Most operating systems still have issues with parallel computation.
  • Most programmers don't know how to write parallel programs.
  • Fewer programmers know how to write efficient parallel programs
  • Fewer still know how to write efficient parallel programs that work correctly.
• Why?
  • Parallel programming introduces a number of problems not faced by sequential programs.
  • Synchronization
  • Scheduling
  • Load Balancing
  • Communications problems
  • Data distribution and access problems.
  • Task decomposition issues
  • Sequential algorithms issues
  • "A supercomputer is a device for converting a CPU-bound problem into an I/O bound problem." [Ken Batcher]