Lock free benchmark results

gistfile1.txt

This document contains information about experiments run to generate the performance data for the engineering blog post about lock-free WiredTiger.

All experiments run on the following hardware:

An Amazon m7g.4xlarge

$ lscpu
Architecture:           aarch64
CPU(s):                 16
Vendor ID:              ARM
  Model:                1
  Thread(s) per core:   1
  Core(s) per socket:   16
  Socket(s):            1
  BogoMIPS:             2100.00


$ cat /proc/meminfo | head -n 1
MemTotal:       64651532 kB

The experiments are configured to run in-memory, so the underlying I/O subsystem isn’t generally important. It is an AWS gp2 drive with 500GB capacity.

WiredTiger was built with the following command line:
cmake -DCMAKE_TOOLCHAIN_FILE=../cmake/toolchains/mongodbtoolchain_v4_gcc.cmake -DCMAKE_INSTALL_PREFIX=$(pwd)/LOCAL_INSTALL -DCC_OPTIMIZE_LEVEL="-O3" -DCMAKE_BUILD_TYPE=Release  -G "Ninja" ./..

From commit c9dfe11d8eec65e13fa1e95291ba1ff3e252bf31

Results were generated by running five times and averaging the result:
./bench/wtperf/wtperf -O ./engblog_writes.wtperf -o verbose=2

Read only
Wtperf:
$ cat build/engblog_writes.wtperf
# wtperf options file: medium btree configuration
conn_config="cache_size=10G,eviction_dirty_trigger=60,eviction_dirty_target=50"
table_config="type=file"
icount=5000000
#random_range=20000000
report_interval=5
run_time=120
populate_threads=1
threads=((count=64,reads=1))
# Add throughput/latency monitoring
max_latency=2000
sample_interval=5

Lock free throughput:
11880941 reads in 5 secs (5 total secs)
12594415 reads in 5 secs (10 total secs)
12563233 reads in 5 secs (15 total secs)
12543354 reads in 5 secs (20 total secs)
12566591 reads in 5 secs (25 total secs)
12573538 reads in 5 secs (30 total secs)
12548016 reads in 5 secs (35 total secs)
12547610 reads in 5 secs (40 total secs)
12551814 reads in 5 secs (45 total secs)
12544169 reads in 5 secs (50 total secs)
12544067 reads in 5 secs (55 total secs)
12555462 reads in 5 secs (60 total secs)
12568297 reads in 5 secs (65 total secs)
12576809 reads in 5 secs (70 total secs)
12586619 reads in 5 secs (75 total secs)
12546612 reads in 5 secs (80 total secs)
12561723 reads in 5 secs (85 total secs)
12587699 reads in 5 secs (90 total secs)
12577511 reads in 5 secs (95 total secs)
12583306 reads in 5 secs (100 total secs)
12583100 reads in 5 secs (105 total secs)
12578665 reads in 5 secs (110 total secs)
12545851 reads in 5 secs (115 total secs)
Executed 300907935 read operations (100%) 2507566 ops/sec

Locked throughput:
8144557 reads in 5 secs (5 total secs)
8547588 reads in 5 secs (10 total secs)
8565919 reads in 5 secs (15 total secs)
8565116 reads in 5 secs (20 total secs)
8559992 reads in 5 secs (25 total secs)
8570315 reads in 5 secs (30 total secs)
8575145 reads in 5 secs (35 total secs)
8571388 reads in 5 secs (40 total secs)
8536120 reads in 5 secs (45 total secs)
8543187 reads in 5 secs (50 total secs)
8544693 reads in 5 secs (55 total secs)
8535100 reads in 5 secs (60 total secs)
8535895 reads in 5 secs (65 total secs)
8561252 reads in 5 secs (70 total secs)
8566152 reads in 5 secs (75 total secs)
8554779 reads in 5 secs (80 total secs)
8555376 reads in 5 secs (85 total secs)
8557400 reads in 5 secs (90 total secs)
8548777 reads in 5 secs (95 total secs)
8543387 reads in 5 secs (100 total secs)
8546087 reads in 5 secs (105 total secs)
8551909 reads in 5 secs (110 total secs)
8551280 reads in 5 secs (115 total secs)
Executed 204867925 read operations (100%) 1707232 ops/sec


Read and insert (append) workload
Wtperf:
$ cat build/engblog_writes.wtperf
# wtperf options file: medium btree configuration
conn_config="cache_size=10G,eviction_dirty_trigger=60,eviction_dirty_target=50"
table_config="type=file"
icount=5000000
random_range=20000000
report_interval=5
run_time=120
populate_threads=1
threads=((count=64,reads=1),(count=6,inserts=1))
# Add throughput/latency monitoring
max_latency=2000
sample_interval=5

Lock free throughput:
747520 inserts, 10329865 reads in 5 secs (5 total secs)
833953 inserts, 10215114 reads in 5 secs (10 total secs)
850921 inserts, 9779752 reads in 5 secs (15 total secs)
767432 inserts, 9632577 reads in 5 secs (20 total secs)
785395 inserts, 9344056 reads in 5 secs (25 total secs)
732163 inserts, 9362662 reads in 5 secs (30 total secs)
730110 inserts, 9153609 reads in 5 secs (35 total secs)
741150 inserts, 9024189 reads in 5 secs (40 total secs)
762275 inserts, 8934668 reads in 5 secs (45 total secs)
719348 inserts, 8951843 reads in 5 secs (50 total secs)
720492 inserts, 8874340 reads in 5 secs (55 total secs)
702197 inserts, 8808282 reads in 5 secs (60 total secs)
627052 inserts, 8802798 reads in 5 secs (65 total secs)
572946 inserts, 8327876 reads in 5 secs (70 total secs)
562412 inserts, 7736815 reads in 5 secs (75 total secs)
563909 inserts, 7872652 reads in 5 secs (80 total secs)
589142 inserts, 8387734 reads in 5 secs (85 total secs)
672068 inserts, 8926825 reads in 5 secs (90 total secs)
888713 inserts, 9253811 reads in 5 secs (95 total secs)
1081388 inserts, 9852593 reads in 5 secs (100 total secs)
757538 inserts, 9744339 reads in 5 secs (105 total secs)
773163 inserts, 9757558 reads in 5 secs (110 total secs)
784126 inserts, 9731641 reads in 5 secs (115 total secs)
Executed 17758564 insert operations (7%) 147988 ops/sec
Executed 220436211 read operations (92%) 1836968 ops/sec

Locked throughput:
492142 inserts, 7281603 reads in 5 secs (5 total secs)
550769 inserts, 7483477 reads in 5 secs (10 total secs)
521246 inserts, 7319455 reads in 5 secs (15 total secs)
546472 inserts, 7164842 reads in 5 secs (20 total secs)
524494 inserts, 7123281 reads in 5 secs (25 total secs)
503132 inserts, 7024998 reads in 5 secs (30 total secs)
510820 inserts, 6939948 reads in 5 secs (35 total secs)
506800 inserts, 6849678 reads in 5 secs (40 total secs)
504194 inserts, 6828193 reads in 5 secs (45 total secs)
499751 inserts, 6801704 reads in 5 secs (50 total secs)
492660 inserts, 6750012 reads in 5 secs (55 total secs)
490302 inserts, 6689944 reads in 5 secs (60 total secs)
486168 inserts, 6680120 reads in 5 secs (65 total secs)
483580 inserts, 6649821 reads in 5 secs (70 total secs)
508498 inserts, 6563780 reads in 5 secs (75 total secs)
505455 inserts, 6566287 reads in 5 secs (80 total secs)
470064 inserts, 6564067 reads in 5 secs (85 total secs)
504409 inserts, 6455171 reads in 5 secs (90 total secs)
441479 inserts, 6419692 reads in 5 secs (95 total secs)
457818 inserts, 6339075 reads in 5 secs (100 total secs)
401442 inserts, 6052977 reads in 5 secs (105 total secs)
367661 inserts, 5850809 reads in 5 secs (110 total secs)
382741 inserts, 5994866 reads in 5 secs (115 total secs)
Executed 11554460 insert operations (6%) 96287 ops/sec
Executed 160448268 read operations (93%) 1337068 ops/sec

Read and Update workload
Wtperf:
$ cat build/engblog_writes.wtperf
# wtperf options file: medium btree configuration
conn_config="cache_size=10G,eviction_dirty_trigger=60,eviction_dirty_target=50"
table_config="type=file"
icount=5000000
#random_range=20000000
report_interval=5
run_time=120
populate_threads=1
threads=((count=64,reads=1),(count=6,updates=1))
# Add throughput/latency monitoring
max_latency=2000
sample_interval=5

Lock free throughput:
9550515 reads, 492715 updates in 5 secs (5 total secs)
10015986 reads, 590684 updates in 5 secs (10 total secs)
10019945 reads, 584407 updates in 5 secs (15 total secs)
10006182 reads, 559782 updates in 5 secs (20 total secs)
10137658 reads, 545949 updates in 5 secs (25 total secs)
9920579 reads, 582482 updates in 5 secs (30 total secs)
9798153 reads, 572807 updates in 5 secs (35 total secs)
9768386 reads, 595908 updates in 5 secs (40 total secs)
9838729 reads, 569578 updates in 5 secs (45 total secs)
9913495 reads, 543309 updates in 5 secs (50 total secs)
9872222 reads, 546800 updates in 5 secs (55 total secs)
9821438 reads, 558135 updates in 5 secs (60 total secs)
9787258 reads, 547445 updates in 5 secs (65 total secs)
9742690 reads, 536274 updates in 5 secs (70 total secs)
9823671 reads, 561804 updates in 5 secs (75 total secs)
9897381 reads, 530832 updates in 5 secs (80 total secs)
9840055 reads, 548205 updates in 5 secs (85 total secs)
9845303 reads, 567630 updates in 5 secs (90 total secs)
9805423 reads, 561814 updates in 5 secs (95 total secs)
9852655 reads, 543259 updates in 5 secs (100 total secs)
9724926 reads, 572016 updates in 5 secs (105 total secs)
9829009 reads, 545018 updates in 5 secs (110 total secs)
9895869 reads, 528959 updates in 5 secs (115 total secs)
Executed 236578646 read operations (94%) 1971488 ops/sec
Executed 13324936 update operations (5%) 111041 ops/sec

Locked throughput:
6892557 reads, 427942 updates in 5 secs (5 total secs)
7240677 reads, 461438 updates in 5 secs (10 total secs)
7263636 reads, 457320 updates in 5 secs (15 total secs)
7275463 reads, 457598 updates in 5 secs (20 total secs)
7263667 reads, 449990 updates in 5 secs (25 total secs)
7211795 reads, 453866 updates in 5 secs (30 total secs)
7239461 reads, 456514 updates in 5 secs (35 total secs)
7224495 reads, 454302 updates in 5 secs (40 total secs)
7208558 reads, 452122 updates in 5 secs (45 total secs)
7188133 reads, 455288 updates in 5 secs (50 total secs)
7144137 reads, 440623 updates in 5 secs (55 total secs)
7198178 reads, 442277 updates in 5 secs (60 total secs)
7186251 reads, 442286 updates in 5 secs (65 total secs)
7161777 reads, 438688 updates in 5 secs (70 total secs)
7153048 reads, 439466 updates in 5 secs (75 total secs)
7125427 reads, 439268 updates in 5 secs (80 total secs)
7206884 reads, 422370 updates in 5 secs (85 total secs)
7189942 reads, 421334 updates in 5 secs (90 total secs)
7202239 reads, 408260 updates in 5 secs (95 total secs)
7185656 reads, 387844 updates in 5 secs (100 total secs)
7204841 reads, 400663 updates in 5 secs (105 total secs)
7266990 reads, 393915 updates in 5 secs (110 total secs)
7265369 reads, 386785 updates in 5 secs (115 total secs)
Executed 172695948 read operations (94%) 1439132 ops/sec
Executed 10387784 update operations (5%) 86564 ops/sec


Multi-threaded append workload
Wtperf:
$ cat build/engblog_writes.wtperf
# wtperf options file: medium btree configuration
conn_config="cache_size=10G,eviction_dirty_trigger=60,eviction_dirty_target=50"
table_config="type=file"
icount=500000
#random_range=20000000
report_interval=5
run_time=120
populate_threads=1
threads=((count=128,inserts=1))
# Add throughput/latency monitoring
max_latency=2000
sample_interval=5

Lock free throughput
9090660 inserts in 5 secs (5 total secs)
8360696 inserts in 5 secs (10 total secs)
7003566 inserts in 5 secs (15 total secs)
5300703 inserts in 5 secs (20 total secs)
7082231 inserts in 5 secs (25 total secs)
5502723 inserts in 5 secs (30 total secs)
6374823 inserts in 5 secs (35 total secs)
4129785 inserts in 5 secs (40 total secs)
5728464 inserts in 5 secs (45 total secs)
7450289 inserts in 5 secs (50 total secs)
6766795 inserts in 5 secs (55 total secs)
6523232 inserts in 5 secs (60 total secs)
4752164 inserts in 5 secs (65 total secs)
6168451 inserts in 5 secs (70 total secs)
6767651 inserts in 5 secs (75 total secs)
6227354 inserts in 5 secs (80 total secs)
7001783 inserts in 5 secs (85 total secs)
5595129 inserts in 5 secs (90 total secs)
6838429 inserts in 5 secs (95 total secs)
5929409 inserts in 5 secs (100 total secs)
6204547 inserts in 5 secs (105 total secs)
5463659 inserts in 5 secs (110 total secs)
4901054 inserts in 5 secs (115 total secs)
Executed 151155647 insert operations (100%) 1259630 ops/sec


Locked throughput
202784 inserts in 5 secs (5 total secs)
1442451 inserts in 5 secs (10 total secs)
1731699 inserts in 5 secs (15 total secs)
1686819 inserts in 5 secs (20 total secs)
1622336 inserts in 5 secs (25 total secs)
1734028 inserts in 5 secs (30 total secs)
1703539 inserts in 5 secs (35 total secs)
1784244 inserts in 5 secs (40 total secs)
1592949 inserts in 5 secs (45 total secs)
1436527 inserts in 5 secs (50 total secs)
1788624 inserts in 5 secs (55 total secs)
1857948 inserts in 5 secs (60 total secs)
1935286 inserts in 5 secs (65 total secs)
1593141 inserts in 5 secs (70 total secs)
1489479 inserts in 5 secs (75 total secs)
1489784 inserts in 5 secs (80 total secs)
1718794 inserts in 5 secs (85 total secs)
1834689 inserts in 5 secs (90 total secs)
1779841 inserts in 5 secs (95 total secs)
1829270 inserts in 5 secs (100 total secs)
1522369 inserts in 5 secs (105 total secs)
1387230 inserts in 5 secs (110 total secs)
1522388 inserts in 5 secs (115 total secs)
Executed 38388402 insert operations (100%) 319903 ops/sec

Changes to add locking to WiredTiger
diff --git a/src/btree/bt_discard.c b/src/btree/bt_discard.c
index 5136e7f99a..d8fe72dd0d 100644
--- a/src/btree/bt_discard.c
+++ b/src/btree/bt_discard.c
@@ -124,6 +124,7 @@ __wt_page_out(WT_SESSION_IMPL *session, WT_PAGE **pagep)
         (void)S2BT(session)->bm->map_discard(
           S2BT(session)->bm, session, dsk, (size_t)dsk->mem_size);


+    __wt_rwlock_destroy(session, &page->tmp_page_rwlock);
     /*
      * If discarding the page as part of process exit, the application may configure to leak the
      * memory rather than do the work.
diff --git a/src/btree/bt_page.c b/src/btree/bt_page.c
index 6768bfc28c..40b1ecefa3 100644
--- a/src/btree/bt_page.c
+++ b/src/btree/bt_page.c
@@ -112,6 +112,7 @@ err:


     /* A new page doesn't have a page id. */
     page->block_meta.page_id = WT_BLOCK_INVALID_PAGE_ID;
+    WT_ERR(__wt_rwlock_init(session, &page->tmp_page_rwlock));


     /* Increment the cache statistics. */
     __wt_cache_page_inmem_incr(session, page, size, false);
diff --git a/src/btree/bt_read.c b/src/btree/bt_read.c
index 9fe24f854c..84b5e4e8a0 100644
--- a/src/btree/bt_read.c
+++ b/src/btree/bt_read.c
@@ -99,6 +99,8 @@ __wt_page_release_evict(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags)


     btree = S2BT(session);


+    /* this function always releases the hazard pointer - also release our read lock */
+    __wt_readunlock(session, &ref->page->tmp_page_rwlock);
     /*
      * This function always releases the hazard pointer - ensure that's done regardless of whether
      * we can get exclusive access. Take some care with order of operations: if we release the
@@ -471,6 +473,7 @@ read:
                 WT_STAT_CONN_INCR(session, page_busy_blocked);
                 break;
             }
+            __wt_readlock(session, &ref->page->tmp_page_rwlock);


             /*
              * If a page has grown too large, we'll try and forcibly evict it before making it
diff --git a/src/btree/bt_sync.c b/src/btree/bt_sync.c
index 0fc30278ce..762dcad2b3 100644
--- a/src/btree/bt_sync.c
+++ b/src/btree/bt_sync.c
@@ -94,8 +94,10 @@ __sync_dup_hazard_pointer(WT_SESSION_IMPL *session, WT_REF *walk)
          * we have one hazard pointer so we should be able to get another one.
          */
         WT_RET(__wt_hazard_set(session, walk, &busy));
-        if (!busy)
+        if (!busy) {
+            __wt_readlock(session, &walk->page->tmp_page_rwlock);
             break;
+        }
         __wt_yield();
     }
     return (0);
diff --git a/src/btree/row_modify.c b/src/btree/row_modify.c
index 6f3d6dd136..96972b938f 100644
--- a/src/btree/row_modify.c
+++ b/src/btree/row_modify.c
@@ -114,8 +114,13 @@ __wt_row_modify(WT_CURSOR_BTREE *cbt, const WT_ITEM *key, const WT_ITEM *value,
           ((value == NULL && upd_arg != NULL) || (value != NULL && upd_arg == NULL))));
     WT_ASSERT(session, upd_arg == NULL || modify_type == WT_UPDATE_INVALID);


+    /* Exclusive indicates a different path than page-read, so the lock isn't held yet */
+    if (!exclusive)
+        __wt_readunlock(session, &page->tmp_page_rwlock);
+    __wt_writelock(session, &page->tmp_page_rwlock);
+
     /* If we don't yet have a modify structure, we'll need one. */
-    WT_RET(__wt_page_modify_init(session, page));
+    WT_ERR(__wt_page_modify_init(session, page));
     mod = page->modify;


     /*
@@ -205,7 +210,8 @@ __wt_row_modify(WT_CURSOR_BTREE *cbt, const WT_ITEM *key, const WT_ITEM *value,
         upd->next = old_upd;


         /* Serialize the update. */
-        WT_ERR(__wt_update_serial(session, cbt, page, upd_entry, &upd, upd_size, exclusive));
+        ret = __wt_update_serial(session, cbt, page, upd_entry, &upd, upd_size, exclusive);
+        WT_ERR(ret);
     } else {
         /*
          * Allocate the insert array as necessary.
@@ -279,8 +285,9 @@ __wt_row_modify(WT_CURSOR_BTREE *cbt, const WT_ITEM *key, const WT_ITEM *value,
                 ins->next[i] = cbt->next_stack[i];


         /* Insert the WT_INSERT structure. */
-        WT_ERR(__wt_insert_serial(
-          session, page, cbt->ins_head, cbt->ins_stack, &ins, ins_size, skipdepth, exclusive));
+        ret = __wt_insert_serial(
+          session, page, cbt->ins_head, cbt->ins_stack, &ins, ins_size, skipdepth, exclusive);
+        WT_ERR(ret);
     }


     inserted_to_update_chain = true;
@@ -345,6 +352,9 @@ err:
         if (upd == NULL && updp_arg != NULL)
             *updp_arg = NULL;
     }
+    __wt_writeunlock(session, &page->tmp_page_rwlock);
+    if (!exclusive)
+        __wt_readlock(session, &page->tmp_page_rwlock);


     return (ret);
 }
diff --git a/src/include/btmem.h b/src/include/btmem.h
index 169c4d431e..ead51b44fa 100644
--- a/src/include/btmem.h
+++ b/src/include/btmem.h
@@ -823,6 +823,7 @@ struct __wt_page {
     uint32_t prefix_start; /* Best page prefix starting slot */
     uint32_t prefix_stop;  /* Maximum slot to which the best page prefix applies */


+    WT_RWLOCK tmp_page_rwlock;
 /* AUTOMATIC FLAG VALUE GENERATION START 0 */
 #define WT_PAGE_BUILD_KEYS 0x001u         /* Keys have been built in memory */
 #define WT_PAGE_COMPACTION_WRITE 0x002u   /* Writing the page for compaction */
diff --git a/src/include/btree_inline.h b/src/include/btree_inline.h
index ea2411d88a..cb4e521833 100644
--- a/src/include/btree_inline.h
+++ b/src/include/btree_inline.h
@@ -2105,6 +2105,7 @@ __wt_page_release(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags)
         }
     }


+    __wt_readunlock(session, &ref->page->tmp_page_rwlock);
     return (__wt_hazard_clear(session, ref));
 }