summaryrefslogtreecommitdiff
path: root/lib/dpif-netlink.c
diff options
context:
space:
mode:
authorMichael Santana <msantana@redhat.com>2022-08-09 03:18:14 -0400
committerIlya Maximets <i.maximets@ovn.org>2022-08-15 18:51:09 +0200
commita5cacea5f988a062f4861c3658252588cf9bf319 (patch)
tree9332dd396c100d69936e3446d6dd1ae35b315d1d /lib/dpif-netlink.c
parent83c9518e7c67fb73ab17f6db50f398dc78403814 (diff)
downloadopenvswitch-a5cacea5f988a062f4861c3658252588cf9bf319.tar.gz
handlers: Create additional handler threads when using CPU isolation.
Additional threads are required to service upcalls when we have CPU isolation (in per-cpu dispatch mode). The reason additional threads are required is because it creates a more fair distribution. With more threads we decrease the load of each thread as more threads would decrease the number of cores each threads is assigned. Adding additional threads also increases the chance OVS utilizes all cores available to use. Some RPS schemas might make some handler threads get all the workload while others get no workload. This tends to happen when the handler thread count is low. An example would be an RPS that sends traffic on all even cores on a system with only the lower half of the cores available for OVS to use. In this example we have as many handlers threads as there are available cores. In this case 50% of the handler threads get all the workload while the other 50% get no workload. Not only that, but OVS is only utilizing half of the cores that it can use. This is the worst case scenario. The ideal scenario is to have as many threads as there are cores - in this case we guarantee that all cores OVS can use are utilized But, adding as many threads are there are cores could have a performance hit when the number of active cores (which all threads have to share) is very low. For this reason we avoid creating as many threads as there are cores and instead meet somewhere in the middle. The formula used to calculate the number of handler threads to create is as follows: handlers_n = min(next_prime(active_cores+1), total_cores) Assume default behavior when total_cores <= 2, that is do not create additional threads when we have less than 2 total cores on the system Fixes: b1e517bd2f81 ("dpif-netlink: Introduce per-cpu upcall dispatch.") Signed-off-by: Michael Santana <msantana@redhat.com> Signed-off-by: Ilya Maximets <i.maximets@ovn.org>
Diffstat (limited to 'lib/dpif-netlink.c')
-rw-r--r--lib/dpif-netlink.c75
1 files changed, 73 insertions, 2 deletions
diff --git a/lib/dpif-netlink.c b/lib/dpif-netlink.c
index 89e1d4325..51eaf7aef 100644
--- a/lib/dpif-netlink.c
+++ b/lib/dpif-netlink.c
@@ -2506,6 +2506,77 @@ dpif_netlink_handler_uninit(struct dpif_handler *handler)
}
#endif
+/* Returns true if num is a prime number,
+ * otherwise, return false.
+ */
+static bool
+is_prime(uint32_t num)
+{
+ if (num == 2) {
+ return true;
+ }
+
+ if (num < 2) {
+ return false;
+ }
+
+ if (num % 2 == 0) {
+ return false;
+ }
+
+ for (uint64_t i = 3; i * i <= num; i += 2) {
+ if (num % i == 0) {
+ return false;
+ }
+ }
+
+ return true;
+}
+
+/* Returns start if start is a prime number. Otherwise returns the next
+ * prime greater than start. Search is limited by UINT32_MAX.
+ *
+ * Returns 0 if no prime has been found between start and UINT32_MAX.
+ */
+static uint32_t
+next_prime(uint32_t start)
+{
+ if (start <= 2) {
+ return 2;
+ }
+
+ for (uint32_t i = start; i < UINT32_MAX; i++) {
+ if (is_prime(i)) {
+ return i;
+ }
+ }
+
+ return 0;
+}
+
+/* Calculates and returns the number of handler threads needed based
+ * the following formula:
+ *
+ * handlers_n = min(next_prime(active_cores + 1), total_cores)
+ */
+static uint32_t
+dpif_netlink_calculate_n_handlers(void)
+{
+ uint32_t total_cores = count_total_cores();
+ uint32_t n_handlers = count_cpu_cores();
+ uint32_t next_prime_num;
+
+ /* If not all cores are available to OVS, create additional handler
+ * threads to ensure more fair distribution of load between them.
+ */
+ if (n_handlers < total_cores && total_cores > 2) {
+ next_prime_num = next_prime(n_handlers + 1);
+ n_handlers = MIN(next_prime_num, total_cores);
+ }
+
+ return n_handlers;
+}
+
static int
dpif_netlink_refresh_handlers_cpu_dispatch(struct dpif_netlink *dpif)
OVS_REQ_WRLOCK(dpif->upcall_lock)
@@ -2515,7 +2586,7 @@ dpif_netlink_refresh_handlers_cpu_dispatch(struct dpif_netlink *dpif)
uint32_t n_handlers;
uint32_t *upcall_pids;
- n_handlers = count_cpu_cores();
+ n_handlers = dpif_netlink_calculate_n_handlers();
if (dpif->n_handlers != n_handlers) {
VLOG_DBG("Dispatch mode(per-cpu): initializing %d handlers",
n_handlers);
@@ -2755,7 +2826,7 @@ dpif_netlink_number_handlers_required(struct dpif *dpif_, uint32_t *n_handlers)
struct dpif_netlink *dpif = dpif_netlink_cast(dpif_);
if (dpif_netlink_upcall_per_cpu(dpif)) {
- *n_handlers = count_cpu_cores();
+ *n_handlers = dpif_netlink_calculate_n_handlers();
return true;
}