Skip to contentSkip to conversation

Fix: hash table growth (for small tables) should be limited

PATCH Mathieu Desnoyers 1 participant 2 messages Aug 30, 2013 2 messages not in the archive

lore
Share & tools
lore
#1 lore raw
In reply to a message not in the archive
Quoted 25 lines · * Mathieu Desnoyers ([email protected]) wrote:

Buckets with many entries encountered in a hash table could cause it to grow to a large size, beyond the scope for which this mechanism is expected to play a role when node accounting is available. Indeed, when the hash table grows to larger size, split-counter node accounting is expected to deal with resize/shrink rather than relying on an heuristic based on the largest bucket size.

This is fixing an issue where we see hash tables sometimes reaching 65k entries index (65536*8 = 524288 bytes) for a workload limited to adding 1000 entries and then removing all of them, done in a loop (random keys).

---

rculfhash.c +0 -3
index 423609d..283cd2d 100644
@@ -718,9 +718,32 @@ void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
 	if (chain_len > 100)
 		dbg_printf("WARNING: large chain length: %u.\n",
 			   chain_len);
-	if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
-		cds_lfht_resize_lazy_grow(ht, size,
-			cds_lfht_get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));

rculfhash.c void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) · line 730

Quoted 10 lines
rculfhash.c +10 -0
+	if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
+		int growth;
+
+		/*
+		 * Ideal growth calculated based on chain length.
+		 */
+		growth = cds_lfht_get_count_order_u32(chain_len
+				- (CHAIN_LEN_TARGET - 1));
+		if ((ht->flags & CDS_LFHT_ACCOUNTING)
+				&& (size << growth) >= (1UL << COUNT_COMMIT_ORDER)) {

Actually, I think it would be cleaner to use the chain_len based mechanism up to

(1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1)

so we end up taking into account that the split-counters are only committed into the global counter every COUNT_COMMIT_ORDER increment/decrement.

It's not bad if we just use (1UL << COUNT_COMMIT_ORDER): we just end up with larger chains in some cases (usually bounded by the log2(number of CPUs)).

Thoughts ?

Thanks,

Mathieu

Quoted 29 lines
rculfhash.c +16 -0
+			/*
+			 * If ideal growth expands the hash table size
+			 * beyond the "small hash table" sizes, use the
+			 * maximum small hash table size to attempt
+			 * expanding the hash table. This only applies
+			 * when node accounting is available, otherwise
+			 * the chain length is used to expand the hash
+			 * table in every case.
+			 */
+			growth = COUNT_COMMIT_ORDER -
+				cds_lfht_get_count_order_u32(size);
+			if (growth <= 0)
+				return;
+		}
+		cds_lfht_resize_lazy_grow(ht, size, growth);
+	}
 }
 
 static
Signature · 9 lines
-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

_______________________________________________
lttng-dev mailing list
[email protected]
http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
#3 lore raw
In reply to a message not in the archive

I'm merging this:

commit 0f9772a234678dd5c775fadab9048f3b9472fe29

Author: Mathieu Desnoyers <[email protected]>
Date:   Tue Aug 27 17:58:22 2013 -0400

Fix: hash table growth (for small tables) should be limited

Buckets with many entries encountered in a hash table could cause it to grow to a large size, beyond the scope for which this mechanism is expected to play a role when node accounting is available. Indeed, when the hash table grows to larger size, split-counter node accounting is expected to deal with resize/shrink rather than relying on an heuristic based on the largest bucket size.

This is fixing an issue where we see hash tables sometimes reaching 65k entries index (65536*8 = 524288 bytes) for a workload limited to adding 1000 entries and then removing all of them, done in a loop (random keys).

rculfhash.c +32 -4
index 423609d..7d39388 100644
@@ -563,6 +563,7 @@ void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
 
 static long nr_cpus_mask = -1;
 static long split_count_mask = -1;
+static int split_count_order = -1;
 
 #if defined(HAVE_SYSCONF)
 static void ht_init_nr_cpus_mask(void)
@@ -597,6 +598,8 @@ void alloc_split_items_count(struct cds_lfht *ht)
 			split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
 		else
 			split_count_mask = nr_cpus_mask;
+		split_count_order =
+			cds_lfht_get_count_order_ulong(split_count_mask + 1);
 	}
 
 	assert(split_count_mask >= 0);
@@ -713,14 +716,39 @@ void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
 	 * Use bucket-local length for small table expand and for
 	 * environments lacking per-cpu data support.
 	 */
-	if (count >= (1UL << COUNT_COMMIT_ORDER))
+	if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
 		return;
 	if (chain_len > 100)
 		dbg_printf("WARNING: large chain length: %u.\n",
 			   chain_len);
-	if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
-		cds_lfht_resize_lazy_grow(ht, size,
-			cds_lfht_get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
+	if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
+		int growth;
+
+		/*
+		 * Ideal growth calculated based on chain length.
+		 */
+		growth = cds_lfht_get_count_order_u32(chain_len
+				- (CHAIN_LEN_TARGET - 1));
+		if ((ht->flags & CDS_LFHT_ACCOUNTING)
+				&& (size << growth)
+					>= (1UL << (COUNT_COMMIT_ORDER
+						+ split_count_order))) {
+			/*
+			 * If ideal growth expands the hash table size
+			 * beyond the "small hash table" sizes, use the
+			 * maximum small hash table size to attempt
+			 * expanding the hash table. This only applies
+			 * when node accounting is available, otherwise
+			 * the chain length is used to expand the hash
+			 * table in every case.
+			 */
+			growth = COUNT_COMMIT_ORDER + split_count_order
+				- cds_lfht_get_count_order_ulong(size);
+			if (growth <= 0)
+				return;
+		}
+		cds_lfht_resize_lazy_grow(ht, size, growth);
+	}
 }
 
 static
Signature · 3 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com