Add code in to prioritize eviction of pages that are larger

than a certain threshold. This avoids taking a performance hit
when a huge page needs to be reconciled.

4 files changed, 40 insertions, 2 deletions

diff --git a/src/btree/bt_evict.c b/src/btree/bt_evict.c
index 02edeabbf47..305ce5ec834 100644
--- a/src/btree/bt_evict.c
+++ b/src/btree/bt_evict.c
@@ -40,9 +40,14 @@ __evict_read_gen(const WT_EVICT_ENTRY *entry)
 {
 	uint64_t read_gen;
 
+	/* Never prioritize empty slots. */
 	if (entry->page == NULL)
 		return (UINT64_MAX);
 
+	/* Always prioritize pages selected by force. */
+	if (__wt_eviction_page_force(entry->btree, entry->page))
+		return (0);
+
 	read_gen = entry->page->read_gen + entry->btree->evict_priority;
 	if (entry->page->type == WT_PAGE_ROW_INT ||
 	    entry->page->type == WT_PAGE_COL_INT)
@@ -279,11 +284,14 @@ __evict_worker(WT_SESSION_IMPL *session)
 		bytes_inuse = __wt_cache_bytes_inuse(cache);
 		dirty_inuse = __wt_cache_dirty_bytes(cache);
 		bytes_max = conn->cache_size;
-		if (bytes_inuse < (cache->eviction_target * bytes_max) / 100 &&
+		if (F_ISSET(cache, WT_EVICT_FORCE_PASS) ||
+		    (bytes_inuse < (cache->eviction_target * bytes_max) / 100 &&
 		    dirty_inuse <
-		    (cache->eviction_dirty_target * bytes_max) / 100)
+		    (cache->eviction_dirty_target * bytes_max) / 100))
 			break;
 
+		F_CLR(cache, WT_EVICT_FORCE_PASS);
+
 		/* Figure out how much we will focus on dirty pages. */
 		if (dirty_inuse >
 		    (cache->eviction_dirty_target * bytes_max) / 100)
diff --git a/src/include/cache.h b/src/include/cache.h
index a70721a5a09..11ddc542411 100644
--- a/src/include/cache.h
+++ b/src/include/cache.h
@@ -73,6 +73,12 @@ struct __wt_cache {
 	uint32_t cp_skip_count;		/* Post change stabilization */
 
 	uint32_t disabled_eviction;	/* Disabled clean / dirty eviction */
+
+	/*
+	 * Flags.
+	 */
+#define	WT_EVICT_FORCE_PASS	0x01	/* Ignore the eviction trigger */
+	uint32_t flags;
 };
 
 /*
diff --git a/src/include/cache.i b/src/include/cache.i
index 73ffe1a4dba..be9cc92dd16 100644
--- a/src/include/cache.i
+++ b/src/include/cache.i
@@ -40,6 +40,19 @@ __wt_eviction_check(WT_SESSION_IMPL *session, int *read_lockoutp, int wake)
 }
 
 /*
+ * __wt_eviction_page_force --
+ *	Return if a page should be a high priority for eviction. The method
+ *      does not use a session handle, since it is called from the eviction
+ *      sort routine, that does not have a session easily available.
+ */
+static inline int
+__wt_eviction_page_force(WT_BTREE *btree, WT_PAGE *page)
+{
+	return (!WT_PAGE_IS_ROOT(page) && __wt_page_is_modified(page) &&
+	    (page->memory_footprint > 20 * btree->maxleafpage));
+}
+
+/*
  * __wt_cache_full_check --
  *	Wait for there to be space in the cache before a read or update.
  */
diff --git a/src/support/hazard.c b/src/support/hazard.c
index 7eb0cf4e5bc..829d0d9cb23 100644
--- a/src/support/hazard.c
+++ b/src/support/hazard.c
@@ -149,6 +149,17 @@ __wt_hazard_clear(WT_SESSION_IMPL *session, WT_PAGE *page)
 	    --hp)
 		if (hp->page == page) {
 			/*
+			 * Trigger eviction if we want to force the page out.
+			 * The page won't necessarily be chosen for eviction 
+			 * every time, but it will be eventually.
+			 */
+			if (__wt_eviction_page_force(btree, page)) {
+				F_SET(S2C(session)->cache,
+				    WT_EVICT_FORCE_PASS);
+				__wt_evict_server_wake(session);
+			}
+
+			/*
 			 * We don't publish the hazard pointer clear in the
 			 * general case.  It's not required for correctness;
 			 * it gives the page server thread faster access to the