Align inter-thread shared structers on cache-line granularity

Specifically ypipe_t specializations for msg_t and command_t.

Currently only supports platforms where posix_memalign is available.
This commit is contained in:
Thomas Rodgers 2015-01-12 13:37:32 -06:00
parent 90194036bf
commit 6372fc2e56
3 changed files with 27 additions and 9 deletions

View File

@ -481,7 +481,7 @@ AM_CONDITIONAL(ON_LINUX, test "x$libzmq_on_linux" = "xyes")
# Checks for library functions.
AC_TYPE_SIGNAL
AC_CHECK_FUNCS(perror gettimeofday clock_gettime memset socket getifaddrs freeifaddrs fork)
AC_CHECK_FUNCS(perror gettimeofday clock_gettime memset socket getifaddrs freeifaddrs fork posix_memalign)
AC_CHECK_HEADERS([alloca.h])
LIBZMQ_CHECK_SOCK_CLOEXEC([

View File

@ -146,6 +146,10 @@ namespace zmq
} done;
} args;
enum { pad_size = 64 - (sizeof(destination) + sizeof(args)) };
unsigned char unused[ pad_size ];
};
}

View File

@ -41,8 +41,16 @@ namespace zmq
// T is the type of the object in the queue.
// N is granularity of the queue (how many pushes have to be done till
// actual memory allocation is required).
#ifdef HAVE_POSIX_MEMALIGN
// ALIGN is the memory alignment size to use in the case where we have
// posix_memalign available. Default value is 64, this alignment will
// prevent two queue chunks from occupying the same CPU cache line on
// architectures where cache lines are <= 64 bytes (e.g. most things
// except POWER).
template <typename T, int N, size_t ALIGN = 64> class yqueue_t
#else
template <typename T, int N> class yqueue_t
#endif
{
public:
@ -103,7 +111,13 @@ namespace zmq
end_chunk->next = sc;
sc->prev = end_chunk;
} else {
#ifdef HAVE_POSIX_MEMALIGN
void *pv;
if (posix_memalign(&pv, ALIGN, sizeof (chunk_t)) == 0)
end_chunk->next = (chunk_t*) pv;
#else
end_chunk->next = (chunk_t*) malloc (sizeof (chunk_t));
#endif
alloc_assert (end_chunk->next);
end_chunk->next->prev = end_chunk;
}