Align all shared memory allocations to cache line boundaries.

Experimentation shows this only costs about 6kB, which seems well
worth it given the major performance effects that can be caused
by insufficient alignment, especially on larger systems.

Discussion: 14166.1458924422@sss.pgh.pa.us

diff --git a/src/backend/storage/ipc/shmem.c b/src/backend/storage/ipc/shmem.c
index 81506eaffd07f65b9cf3cfa395964c2bfe92f6ed..1ad68cda4e517983f0837083197d2d312ce90c4a 100644 (file)
--- a/src/backend/storage/ipc/shmem.c
+++ b/src/backend/storage/ipc/shmem.c
@@ -171,9 +171,17 @@ ShmemAlloc(Size size)
        void       *newSpace;
 
        /*
-        * ensure all space is adequately aligned.
+        * Ensure all space is adequately aligned.  We used to only MAXALIGN this
+        * space but experience has proved that on modern systems that is not good
+        * enough.  Many parts of the system are very sensitive to critical data
+        * structures getting split across cache line boundaries.  To avoid that,
+        * attempt to align the beginning of the allocation to a cache line
+        * boundary.  The calling code will still need to be careful about how it
+        * uses the allocated space - e.g. by padding each element in an array of
+        * structures out to a power-of-two size - but without this, even that
+        * won't be sufficient.
         */
-       size = MAXALIGN(size);
+       size = CACHELINEALIGN(size);
 
        Assert(ShmemSegHdr != NULL);