Improve pair search thread load balance

With very small systems and many OpenMP threads, especially when
using GPUs, some threads can end up without pair search work. Better
load balancing reduces the pair search time. Also the CPU non-bonded
kernel time is slightly reduced in the extreme parallelization limit.

Change-Id: Ib036ea3ba59f497eeee7afa73a71fb0e0ccd216e

diff --git a/src/gromacs/mdlib/nbnxn_search.c b/src/gromacs/mdlib/nbnxn_search.c
index 7e278a94ee814f077610956989ae9eb36e5fc703..6ec71ccd2a2e38e5fa25c82c7dcc08794a94fbb1 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_search.c
+++ b/src/gromacs/mdlib/nbnxn_search.c
@@ -4818,11 +4818,24 @@ static int get_ci_block_size(const nbnxn_grid_t *gridi,
     /* Without domain decomposition
      * or with less than 3 blocks per task, divide in nth blocks.
      */
-    if (!bDomDec || ci_block*3*nth > gridi->nc)
+    if (!bDomDec || nth*3*ci_block > gridi->nc)
     {
         ci_block = (gridi->nc + nth - 1)/nth;
     }
 
+    if (ci_block > 1 && (nth - 1)*ci_block >= gridi->nc)
+    {
+        /* Some threads have no work. Although reducing the block size
+         * does not decrease the block count on the first few threads,
+         * with GPUs better mixing of "upper" cells that have more empty
+         * clusters results in a somewhat lower max load over all threads.
+         * Without GPUs the regime of so few atoms per thread is less
+         * performance relevant, but with 8-wide SIMD the same reasoning
+         * applies, since the pair list uses 4 i-atom "sub-clusters".
+         */
+        ci_block--;
+    }
+
     return ci_block;
 }