LogP machine

From HandWiki

The LogP machine is a model for parallel computation.[1] It aims at being more practical than the PRAM model while still allowing for easy analysis of computation. The name is not related to the mathematical logarithmic function: Instead, the machine is described by the four parameters [math]\displaystyle{ L }[/math], [math]\displaystyle{ o }[/math], [math]\displaystyle{ g }[/math] and [math]\displaystyle{ P }[/math].

The LogP machine consists of arbitrarily many processing units with distributed memory. The processing units are connected through an abstract communication medium which allows point-to-point communication. This model is pair-wise synchronous and overall asynchronous.

The machine is described by the four parameters:

  • [math]\displaystyle{ L }[/math], the latency of the communication medium.
  • [math]\displaystyle{ o }[/math], the overhead of sending and receiving a message.
  • [math]\displaystyle{ g }[/math], the gap required between two send/receive operations. A more common interpretation of this quantity is as the inverse of the bandwidth of a processor-processor communication channel.
  • [math]\displaystyle{ P }[/math], the number of processing units.

Each local operation on each machine takes the same time ('unit time'). This time is called a processor cycle. The units of the parameters [math]\displaystyle{ L }[/math], [math]\displaystyle{ o }[/math] and [math]\displaystyle{ g }[/math] are measured in multiples of processor cycles.

See also

Notes

  1. Culler et al. 1993

References

Culler, David; Karp, Richard; Patterson, David; Sahay, Abhijit; Schauser, Klaus Erik; Santos; Subramonian, Ramesh; Von Eicken, Thorsten (July 1993), "LogP: Towards a realistic model of parallel computation", ACM SIGPLAN Notices 28 (7): 1–12, doi:10.1145/173284.155333, https://dl.acm.org/doi/pdf/10.1145/173284.155333