LoBoS benchmarks, Jan/Feb 2000, c28n1

Cubic periodic molecular systems, OG = octyl glucoside:

  • M100 aqueous micelle of 100 OG lipids, ca. 19K atoms
  • M50 aqueous micelle of 50 OG lipids, ca. 17K atoms
  • M20 aqueous micelle of 20 OG lipids, ca. 8K atoms
  • bx30 aqueous sucrose, 30 A box edge, ca. 3K atoms

    NEW! 5 Feb 2000 Results for 2.2.14-tcpfix

    Plots are all PM Ewald tests, mostly GbNIC; spherical cutoff results are typically slightly better, except for larger systems on 100baseT. In that case, PM Ewald and spherical cutoff results are approximately equivalent.

    Note that while systems of ca. 3K atoms (bx30) show little benefit at 8 processors, systems of 8K (M20) and up all show a ca. 25% reduction in elapsed time, compared to 4 processors. The relatively small gap between 2 and 1 proc/host seen at 8 processors is the smallest we've ever observed on this hardware. The M100 results for the 2.2.13 kernel (OLD) are shown to illustrate the change.


    The gap between curves is significant in the above plot; it shows: (a) that there's always an advantage to using 4 procs, even for 3K atoms, (b) that systems over about 4K atoms would benefit from using 8 procs, (c) the small gap between 2 and 1 proc/host, and (d) GbNIC only offers any real advantage for large systems using 8 procs. These are all significant improvements in parallel scaling performance, especially for the 2 proc/host case. The GbNIC 8 proc results for the 2.2.13 kernel (OLD) are shown to illustrate the change.


    The above shows the 8 proc results for optimizing the kernel parameter tcp_delack_strategy on Gigabit ethernet (GbNIC); note the near convergence of 2 and 1 proc/host curves, such that at 0 and 1 the times are ranked by molecular system size.


    As for the previous plot, but for fast ethernet (100baseT).

    Credits