commit df7b7bf64eb354114e6c519e3e03ffc446afa8ba
Author: Nathan Scott <nathans@redhat.com>
Date: Fri Nov 26 09:17:23 2021 +1100
libpcp_pmda: add indom cache fast-paths for inst lookup beyond max
We encountered a situation where indom cache loading consumed vast
CPU resources for an indom of size ~150k instances. Profiling was
used to identify the insert loop that ensures the inst linked list
within the cache hash tables is sorted - this loop is O(N*2) as we
potentially walk this list from the start on every insert during a
cache load. Because cache loading happens from a sorted file, the
worst-case scenario happened every time - each new instance insert
occurs beyond the current maximum. Fortunately we maintain a last
entry pointer, so the new fast path uses that first and falls back
to the original behaviour for an out-of-order insertion.
A second opportunity for the same optimization was identified when
auditing the rest of cache.c - in the find_inst() routine for inst
identifier lookups beyond the current maximum observed instance.
Resolves Red Hat BZ #2024648
diff --git a/src/libpcp_pmda/src/cache.c b/src/libpcp_pmda/src/cache.c
index 0e66506d74..196ffc1da9 100644
--- a/src/libpcp_pmda/src/cache.c
+++ b/src/libpcp_pmda/src/cache.c
@@ -328,6 +328,9 @@ find_inst(hdr_t *h, int inst)
{
entry_t *e;
+ if ((e = h->last) != NULL && e->inst < inst)
+ return NULL;
+
for (e = h->first; e != NULL; e = e->next) {
if (e->inst == inst && e->state != PMDA_CACHE_EMPTY)
break;
@@ -621,7 +624,11 @@ insert_cache(hdr_t *h, const char *name, int inst, int *sts)
*sts = PM_ERR_INST;
return e;
}
- for (e = h->first; e != NULL; e = e->next) {
+ /* if this entry is beyond the (sorted) list end, avoid linear scan */
+ if ((e = h->last) == NULL || e->inst > inst)
+ e = h->first;
+ /* linear search over linked list, starting at either first or last */
+ for (; e != NULL; e = e->next) {
if (e->inst < inst)
last_e = e;
else if (e->inst > inst)