Multi Range Read optimization

Multi Range Read is an optimization aimed at improving performance for IO-bound queries that need to scan lots of rows.

Multi Range Read is available in MariaDB since MariaDB 5.3, and MySQL has a subset of it in MySQL 5.6.

Multi Range Read can be used with

• range access
• ref and eq_ref access, when they are using Batched Key Access

as shown in this diagram:

The idea

Case 1: rowid sorting for range access

Consider a range query:

MariaDB [test]> explain select * from tbl where tbl.key1 between 1000 and 2000;
+----+-------------+-------+-------+---------------+------+---------+------+------+-----------------------+
| id | select_type | table | type  | possible_keys | key  | key_len | ref  | rows | Extra                 |
+----+-------------+-------+-------+---------------+------+---------+------+------+-----------------------+
|  1 | SIMPLE      | tbl   | range | key1          | key1 | 5       | NULL |  960 | Using index condition |
+----+-------------+-------+-------+---------------+------+---------+------+------+-----------------------+

When this query is executed, disk IO access pattern will follow the red line in this figure:

Execution will hit the table rows in random places, as marked with the blue line/numbers in the figure.

When the table is sufficiently big, each table record read will need to actually go to disk (and be served from buffer pool or OS cache), and query execution will be too slow to be practical. For example, a 10,000 RPM disk drive is able to make 167 seeks per second, so in the worst case, query execution will be capped at reading about 167 records per second.

SSD drives do not need to do disk seeks, so they will not be hurt as badly, however the performance will still be poor in many cases.

Multi-Range-Read optimization aims to make disk access faster by sorting record read requests and then doing one ordered disk sweep. If one enables Multi Range Read, EXPLAIN will show that a "Rowid-ordered scan" is used:

MariaDB [test]> set optimizer_switch='mrr=on';
Query OK, 0 rows affected (0.06 sec)

MariaDB [test]> explain select * from tbl where tbl.key1 between 1000 and 2000;
+----+-------------+-------+-------+---------------+------+---------+------+------+-------------------------------------------+
| id | select_type | table | type  | possible_keys | key  | key_len | ref  | rows | Extra                                     |
+----+-------------+-------+-------+---------------+------+---------+------+------+-------------------------------------------+
|  1 | SIMPLE      | tbl   | range | key1          | key1 | 5       | NULL |  960 | Using index condition; Rowid-ordered scan |
+----+-------------+-------+-------+---------------+------+---------+------+------+-------------------------------------------+
1 row in set (0.03 sec)

and the execution will proceed as follows:

Reading disk data sequentially is generally faster, because

• Rotating drives do not have to move the head back and forth
• One can take advantage of IO-prefetching done at various levels
• Each disk page will be read exactly once, which means we won't rely on disk cache (or buffer pool) to save us from reading the same page multiple times.

The above can make a huge difference on performance. There is also a catch, though:

• If you're scanning small data ranges in a table that is sufficiently small so that it completely fits into the OS disk cache, then you may observe that the only effect of MRR is that extra buffering/sorting adds some CPU overhead.

• LIMIT n and ORDER BY ... LIMIT n queries with small values of n may become slower. The reason is that MRR reads data in disk order, while ORDER BY ... LIMIT n wants first n records in index order.

Case 2: Rowid sorting for Batched Key Access

Batched Key Access can benefit from rowid sorting in the same way as range access does. If one has a join that uses index lookups:

MariaDB [test]> explain select * from t1,t2 where t2.key1=t1.col1;
+----+-------------+-------+------+---------------+------+---------+--------------+------+-------------+
| id | select_type | table | type | possible_keys | key  | key_len | ref          | rows | Extra       |
+----+-------------+-------+------+---------------+------+---------+--------------+------+-------------+
|  1 | SIMPLE      | t1    | ALL  | NULL          | NULL | NULL    | NULL         | 1000 | Using where |
|  1 | SIMPLE      | t2    | ref  | key1          | key1 | 5       | test.t1.col1 |    1 |             |
+----+-------------+-------+------+---------------+------+---------+--------------+------+-------------+
2 rows in set (0.00 sec)

Execution of this query will cause table t2 to be hit in random locations by lookups made through t2.key1=t1.col. If you enable Multi Range and and Batched Key Access, you will get table t2 to be accessed using a Rowid-ordered scan:

MariaDB [test]> set optimizer_switch='mrr=on';
Query OK, 0 rows affected (0.06 sec)

MariaDB [test]> set join_cache_level=6;
Query OK, 0 rows affected (0.00 sec)

MariaDB [test]> explain select * from t1,t2 where t2.key1=t1.col1;
+----+-------------+-------+------+---------------+------+---------+--------------+------+--------------------------------------------------------+
| id | select_type | table | type | possible_keys | key  | key_len | ref          | rows | Extra                                                  |
+----+-------------+-------+------+---------------+------+---------+--------------+------+--------------------------------------------------------+
|  1 | SIMPLE      | t1    | ALL  | NULL          | NULL | NULL    | NULL         | 1000 | Using where                                            |
|  1 | SIMPLE      | t2    | ref  | key1          | key1 | 5       | test.t1.col1 |    1 | Using join buffer (flat, BKA join); Rowid-ordered scan |
+----+-------------+-------+------+---------------+------+---------+--------------+------+--------------------------------------------------------+
2 rows in set (0.00 sec)

The benefits will be similar to those listed for range access.

An additional source of speedup is this property: if there are multiple records in t1 that have the same value of t1.col1, then regular Nested-Loops join will make multiple index lookups for the same value of t2.key1=t1.col1. The lookups may or may not hit the cache, depending on how big the join is. With Batched Key Access and Multi-Range Read, no duplicate index lookups will be made.

Case 3: Key sorting for Batched Key Access

Let us consider again example of neted loop join, with ref access on the second table:

MariaDB [test]> explain select * from t1,t2 where t2.key1=t1.col1;
+----+-------------+-------+------+---------------+------+---------+--------------+------+-------------+
| id | select_type | table | type | possible_keys | key  | key_len | ref          | rows | Extra       |
+----+-------------+-------+------+---------------+------+---------+--------------+------+-------------+
|  1 | SIMPLE      | t1    | ALL  | NULL          | NULL | NULL    | NULL         | 1000 | Using where |
|  1 | SIMPLE      | t2    | ref  | key1          | key1 | 5       | test.t1.col1 |    1 |             |
+----+-------------+-------+------+---------------+------+---------+--------------+------+-------------+

Execution of this query plan will cause random hits to be made into the index t2.key1, as shown in this picture:

In particular, on step #5 we'll read the same index page that we've read on step #2, and the page we've read on step #4 will be re-read on step#6. If all pages you're accessing are in the cache (in the buffer pool, if you're using InnoDB, and in the key cache, if you're using MyISAM), this is not a problem. However, if your hit ratio is poor and you're going to hit the disk, it makes sense to sort the lookup keys, like shown in this figure:

This is roughly what Key-ordered scan optimization does. In EXPLAIN, it looks like follows:

MariaDB [test]> set optimizer_switch='mrr=on,mrr_sort_keys=on';
Query OK, 0 rows affected (0.00 sec)

MariaDB [test]> set join_cache_level=6;
Query OK, 0 rows affected (0.02 sec)

MariaDB [test]> explain select * from t1,t2 where t2.key1=t1.col1\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
         type: ALL
possible_keys: a
          key: NULL
      key_len: NULL
          ref: NULL
         rows: 1000
        Extra: Using where
*************************** 2. row ***************************
           id: 1
  select_type: SIMPLE
        table: t2
         type: ref
possible_keys: key1
          key: key1
      key_len: 5
          ref: test.t1.col1
         rows: 1
        Extra: Using join buffer (flat, BKA join); Key-ordered Rowid-ordered scan
2 rows in set (0.00 sec)

((TODO: a note about why sweep-read over InnoDB's clustered primary index scan (which is, actually the whole InnoDB table itself) will use Key-ordered scan algorithm, but not Rowid-ordered scan algorithm, even though conceptually they are the same thing in this case))

Buffer space management

As was shown above, Multi Range Read requires sort buffers to operate. The size of the buffers is limited by system variables. If MRR has to process more data than it can fit into its buffer, it will break the scan into multiple passes. The more passes are made, the less is the speedup though, so one needs to balance between having too big buffers (which consume lots of memory) and too small buffers (which limit the possible speedup).

Range access

When MRR is used for range access, the size of its buffer is controlled by the @mrr_buffer_size system variable. Its value specifies how much space can be used for each table. For example, if there is a query which is a 10-way join and MRR is used for each table, 10*@@mrr_buffer_size bytes may be used.

Batched Key Access

When Multi Range Read is used by Batched Key Access, then buffer space is managed by BKA code, which will automatically provide a part of its buffer space to MRR. You can control the amount of space used by BKA by setting

• @@join_buffer_size to limit how much memory BKA uses for each table, and
• @@join_cache_space_limit to limit the total amount of memory used by BKA in the join.

Counters

Multi Range Read has three counters

• Handler_mrr_init - counts how many MRR scans were performed
• Handler_mrr_extra_rowid_sorts - counts rowid sort operations done in MRR scans, except the first scan in each operation
• Handler_mrr_extra_key_sorts - counts key sort operations done in MRR scans, except the first scan in each operation

One possible use for Handler_mrr_extra_rowid_sorts and Handler_mrr_extra_key_sorts values: if they are non-zero, it means that MRR had to make multiple passes, and you may want to increase buffer sizes.

An important thing to note: When MRR scan makes an index lookup (or other "basic" operation), it will cause its "basic" counter, like Handler_read_key to be incremented, also. This way, you can still see total number of index accesses, including those made by MRR. Per-user/table/index statistics also include operations made by MRR.

Multi Range Read Factsheet

• Multi Range Read is used by
  • range access method for range scans.
  • Batched Key Access for joins
• Multi Range Read can cause slowdowns for small queries over small tables, so it is disabled by default.
• There are two strategies:
  • Rowid-ordered scan
  • Key-ordered scan
• : and you can tell if either of them is used by checking the Extra column in EXPLAIN output.
• There are three @@optimizer_switch flags you can switch ON:
  • mrr=on - enable MRR and rowid ordered scans
  • mrr_sort_keys=on - enable Key-ordered scans (you must also set mrr=on for this to have any effect)
  • mrr_cost_based=on - enable cost-based choice whether to use MRR. Currently not recommended, because cost model is not sufficiently tuned yet.

Differences from MySQL

• MySQL supports only Rowid ordered scan strategy, which it shows in EXPLAIN as Using MRR.
• EXPLAIN in MySQL shows Using MRR, while in MariaDB it may show
  • Rowid-ordered scan
  • Key-ordered scan
  • Key-ordered Rowid-ordered scan
• MariaDB uses @@mrr_buffer_size as a limit of MRR buffer size for range access, while MySQL uses @@read_rnd_buffer_size.
• MariaDB has three MRR counters: Handler_mrr_init, Handler_mrr_extra_rowid_sorts, Handler_mrr_extra_key_sorts, while MySQL has only Handler_mrr_init, and it will only count MRR scans that were used by BKA. MRR scans used by range access are not counted.

See also

• What is MariaDB 5.3
• Multi-Range Read Optimization page in MySQL manual