Dirty buffer throttling v16

Suggest edits

EDB Resource Manager uses dirty buffer throttling to keep the aggregate shared buffer writing rate of all processes in the group near the limit specified by the dirty_rate_limit parameter. A process in the group might be interrupted and put into sleep mode for a short time to maintain the defined limit. When and how such interruptions occur is defined by a proprietary algorithm used by EDB Resource Manager.

To control writing to shared buffers, set the dirty_rate_limit resource type parameter.

  • Set the dirty_rate_limit parameter to the number of kilobytes per second for the combined rate at which all the processes in the group write to, or “dirty”, the shared buffers. An example setting is 3072 kilobytes per seconds.

  • The valid range of the dirty_rate_limit parameter is 0 to 1.67772e+07. A setting of 0 means no dirty rate limit was set for the resource group.

Setting the dirty rate limit for a resource group

Use the ALTER RESOURCE GROUP command with the SET dirty_rate_limit clause to set the dirty rate limit for a resource group.

In this example, the dirty rate limit is set to 12288 kilobytes per second for resgrp_a, 6144 kilobytes per second for resgrp_b, and 3072 kilobytes per second for resgrp_c. This means that the combined writing rate to the shared buffer of all processes assigned to resgrp_a is maintained at approximately 12288 kilobytes per second. Similarly, for all processes in resgrp_b, the combined writing rate to the shared buffer is kept to approximately 6144 kilobytes per second, and so on.

edb=# ALTER RESOURCE GROUP resgrp_a SET dirty_rate_limit TO 12288;
ALTER RESOURCE GROUP
edb=# ALTER RESOURCE GROUP resgrp_b SET dirty_rate_limit TO 6144;
ALTER RESOURCE GROUP
edb=# ALTER RESOURCE GROUP resgrp_c SET dirty_rate_limit TO 3072;
ALTER RESOURCE GROUP

This query shows the settings of dirty_rate_limit in the catalog;

edb=# SELECT rgrpname, rgrpdirtyratelimit FROM edb_resource_group;
Output
 rgrpname  | rgrpdirtyratelimit
-----------+--------------------
 resgrp_a  |         12288
 resgrp_b  |         6144
 resgrp_c  |         3072
(3 rows)

Changing the dirty rate limit

Changing the dirty_rate_limit of a resource group affects new processes that are assigned to the group. Any currently running processes that are members of the group are also immediately affected by the change. That is, if the dirty_rate_limit is changed from 12288 to 3072, currently running processes in the group are throttled downward so that the aggregate group dirty rate is near 3072 kilobytes per second instead of 12288 kilobytes per second.

To show the effect of setting the dirty rate limit for resource groups, the examples use the following table for intensive I/O operations:

CREATE TABLE t1 (c1 INTEGER, c2 CHARACTER(500)) WITH (FILLFACTOR = 10);

The FILLFACTOR = 10 clause results in INSERT commands packing rows up to only 10% per page. The result is a larger sampling of dirty shared blocks for the purpose of these examples.

Displaying the number of dirty buffers

The pg_stat_statements module is used to display the number of shared buffer blocks that are dirtied by a SQL command and the amount of time the command took to execute. This information is used to calculate the actual kilobytes per second writing rate for the SQL command and thus compare it to the dirty rate limit set for a resource group.

To use the pg_stat_statements module:

  1. In the postgresql.conf file, add $libdir/pg_stat_statements to the shared_preload_libraries configuration parameter:

    shared_preload_libraries = '$libdir/dbms_pipe,$libdir/edb_gen,$libdir/pg_stat_statements'

  2. Restart the database server.

  3. Use the CREATE EXTENSION command to finish creating the pg_stat_statements module:

    edb=# CREATE EXTENSION pg_stat_statements SCHEMA public;
CREATE EXTENSION

    The pg_stat_statements_reset() function clears out the pg_stat_statements view for clarity of each example.

The resource groups with the dirty rate limit settings shown in the previous query are used in these examples.

Example: Single process in a single group

This sequence of commands creates table t1. The current process is set to use resource group resgrp_b. The pg_stat_statements view is cleared out by running the pg_stat_statements_reset() function.

The INSERT command then generates a series of integers from 1 to 10,000 to populate the table and dirty approximately 10,000 blocks:

edb=# CREATE TABLE t1 (c1 INTEGER, c2 CHARACTER(500)) WITH (FILLFACTOR = 10);
CREATE TABLE
edb=# SET edb_resource_group TO resgrp_b;
SET
edb=# SHOW edb_resource_group;
Output
 edb_resource_group
--------------------
 resgrp_b
(1 row)
edb=# SELECT pg_stat_statements_reset();
Output
 pg_stat_statements_reset
--------------------------

(1 row)

edb=# INSERT INTO t1 VALUES (generate_series (1,10000), 'aaa');
INSERT 0 10000

This example shows the results from the INSERT command:

edb=# SELECT query, rows, total_time, shared_blks_dirtied FROM
pg_stat_statements;
Output
-[ RECORD 1 ]--------+--------------------------------------------------
 query               | INSERT INTO t1 VALUES (generate_series (?,?), ?);
 rows                | 10000
 total_time          | 13496.184
 shared_blks_dirtied | 10003

The actual dirty rate is calculated as follows:

  • The number of blocks dirtied per millisecond (ms) is 10003 blocks / 13496.184 ms, which yields 0.74117247 blocks per millisecond.
  • Multiply the result by 1000 to give the number of shared blocks dirtied per second (1 second = 1000 ms), which yields 741.17247 blocks per second.
  • Multiply the result by 8.192 to give the number of kilobytes dirtied per second (1 block = 8.192 kilobytes), which yields approximately 6072 kilobytes per second.

The actual dirty rate of 6072 kilobytes per second is close to the dirty rate limit for the resource group, which is 6144 kilobytes per second.

By contrast, if you repeat the steps without the process belonging to any resource group, the dirty buffer rate is much higher:

edb=# CREATE TABLE t1 (c1 INTEGER, c2 CHARACTER(500)) WITH (FILLFACTOR = 10);
CREATE TABLE
edb=# SHOW edb_resource_group;
Output
 edb_resource_group
--------------------

(1 row)
edb=# SELECT pg_stat_statements_reset();
Output
 pg_stat_statements_reset
--------------------------

(1 row)

edb=# INSERT INTO t1 VALUES (generate_series (1,10000), 'aaa');
INSERT 0 10000

This example shows the results from the INSERT command without the use of a resource group:

edb=# SELECT query, rows, total_time, shared_blks_dirtied FROM
pg_stat_statements;
Output
-[ RECORD 1 ]--------+--------------------------------------------------
 query               | INSERT INTO t1 VALUES (generate_series (?,?), ?);
 rows                | 10000
 total_time          | 2432.165
 shared_blks_dirtied | 10003

The total time was only 2432.165 milliseconds, compared to 13496.184 milliseconds when using a resource group with a dirty rate limit set to 6144 kilobytes per second.

The actual dirty rate without the use of a resource group is calculated as follows:

  • The number of blocks dirtied per millisecond (ms) is 10003 blocks / 2432.165 ms, which yields 4.112797 blocks per millisecond.
  • Multiply the result by 1000 to give the number of shared blocks dirtied per second (1 second = 1000 ms), which yields 4112.797 blocks per second.
  • Multiply the result by 8.192 to give the number of kilobytes dirtied per second (1 block = 8.192 kilobytes), which yields approximately 33692 kilobytes per second.

The actual dirty rate of 33692 kilobytes per second is much higher than when the resource group with a dirty rate limit of 6144 kilobytes per second was used.

Example: Multiple processes in a single group

As stated previously, the dirty rate limit applies to the aggregate of all processes in the resource group. This concept is illustrated in the following example.

For this example, the inserts are performed simultaneously on two different tables in two separate psql sessions, each of which was added to resource group resgrp_b that has a dirty_rate_limit set to 6144 kilobytes per second.

Session 1

edb=# CREATE TABLE t1 (c1 INTEGER, c2 CHARACTER(500)) WITH (FILLFACTOR = 10);
CREATE TABLE
edb=# SET edb_resource_group TO resgrp_b;
SET
edb=# SHOW edb_resource_group;
Output
 edb_resource_group
--------------------
 resgrp_b
(1 row)

edb=# INSERT INTO t1 VALUES (generate_series (1,10000), 'aaa');
INSERT 0 10000

Session 2

edb=# CREATE TABLE t2 (c1 INTEGER, c2 CHARACTER(500)) WITH (FILLFACTOR = 10);
CREATE TABLE
edb=# SET edb_resource_group TO resgrp_b;
SET
edb=# SHOW edb_resource_group;
Output
 edb_resource_group
--------------------
 resgrp_b
(1 row)
edb=# SELECT pg_stat_statements_reset();
Output
 pg_stat_statements_reset
--------------------------
(1 row)

edb=# INSERT INTO t2 VALUES (generate_series (1,10000), 'aaa');
INSERT 0 10000
Note

The INSERT commands in session 1 and session 2 started after the SELECT pg_stat_statements_reset() command in session 2 ran.

This example shows the results from the INSERT commands in the two sessions. RECORD 3 shows the results from session 1. RECORD 2 shows the results from session 2.

edb=# SELECT query, rows, total_time, shared_blks_dirtied FROM
pg_stat_statements;
Output
-[ RECORD 1 ]--------+--------------------------------------------------
 query               | SELECT pg_stat_statements_reset();
 rows                | 1
 total_time          | 0.43
 shared_blks_dirtied | 0
-[ RECORD 2 ]--------+--------------------------------------------------
 query               | INSERT INTO t2 VALUES (generate_series (?,?), ?);
 rows                | 10000
 total_time          | 30591.551
 shared_blks_dirtied | 10003
-[ RECORD 3 ]--------+--------------------------------------------------
 query               | INSERT INTO t1 VALUES (generate_series (?,?), ?);
 rows                | 10000
 total_time          | 33215.334
 shared_blks_dirtied | 10003

The total time was 33215.334 milliseconds for session 1 and 30591.551 milliseconds for session 2. When only one session was active in the same resource group, the time was 13496.184 milliseconds. Thus, more active processes in the resource group result in a slower dirty rate for each active process in the group. The following calculations show this.

The actual dirty rate for session 1 is calculated as follows:

  • The number of blocks dirtied per millisecond (ms) is 10003 blocks / 33215.334 ms, which yields 0.30115609 blocks per millisecond.
  • Multiply the result by 1000 to give the number of shared blocks dirtied per second (1 second = 1000 ms), which yields 301.15609 blocks per second.
  • Multiply the result by 8.192 to give the number of kilobytes dirtied per second (1 block = 8.192 kilobytes), which yields approximately 2467 kilobytes per second.

The actual dirty rate for session 2 is calculated as follows:

  • The number of blocks dirtied per millisecond (ms) is 10003 blocks / 30591.551 ms, which yields 0.32698571 blocks per millisecond.
  • Multiply the result by 1000 to give the number of shared blocks dirtied per second (1 second = 1000 ms), which yields 326.98571 blocks per second.
  • Multiply the result by 8.192 to give the number of kilobytes dirtied per second (1 block = 8.192 kilobytes), which yields approximately 2679 kilobytes per second.

The combined dirty rate from session 1 (2467 kilobytes per second) and from session 2 (2679 kilobytes per second) yields 5146 kilobytes per second, which is below the set dirty rate limit of the resource group (6144 kilobytes per seconds).

Example: Multiple processes in multiple groups

In this example, two additional psql sessions are used along with the previous two sessions. The third and fourth sessions perform the same INSERT command in resource group resgrp_c with a dirty_rate_limit of 3072 kilobytes per second.

Repeat sessions 1 and 2 from the prior example using resource group resgrp_b with a dirty_rate_limit of 6144 kilobytes per second:

Session 3

edb=# CREATE TABLE t3 (c1 INTEGER, c2 CHARACTER(500)) WITH (FILLFACTOR = 10);
CREATE TABLE
edb=# SET edb_resource_group TO resgrp_c;
SET
edb=# SHOW edb_resource_group;
Output
 edb_resource_group
--------------------


resgrp_c
(1 row)

edb=# INSERT INTO t3 VALUES (generate_series (1,10000), 'aaa');
INSERT 0 10000

Session 4

edb=# CREATE TABLE t4 (c1 INTEGER, c2 CHARACTER(500)) WITH (FILLFACTOR = 10);
CREATE TABLE
edb=# SET edb_resource_group TO resgrp_c;
SET
edb=# SHOW edb_resource_group;
Output
 edb_resource_group
--------------------
 resgrp_c
(1 row)
edb=# SELECT pg_stat_statements_reset();
Output
 pg_stat_statements_reset
--------------------------

(1 row)

edb=# INSERT INTO t4 VALUES (generate_series (1,10000), 'aaa');
INSERT 0 10000
Note

The INSERT commands in all four sessions started after the SELECT pg_stat_statements_reset() command in session 4 ran.

This example shows the results from the INSERT commands in the four sessions:

  • RECORD 3 shows the results from session 1. RECORD 2 shows the results from session 2.

  • RECORD 4 shows the results from session 3. RECORD 5 shows the results from session 4.

edb=# SELECT query, rows, total_time, shared_blks_dirtied FROM
pg_stat_statements;
Output
-[ RECORD 1 ]--------+--------------------------------------------------
 query               | SELECT pg_stat_statements_reset();
 rows                | 1
 total_time          | 0.467
 shared_blks_dirtied | 0
-[ RECORD 2 ]--------+--------------------------------------------------
 query               | INSERT INTO t2 VALUES (generate_series (?,?), ?);
 rows                | 10000
 total_time          | 31343.458
 shared_blks_dirtied | 10003
-[ RECORD 3 ]--------+--------------------------------------------------
 query               | INSERT INTO t1 VALUES (generate_series (?,?), ?);
 rows                | 10000
 total_time          | 28407.435
 shared_blks_dirtied | 10003
-[ RECORD 4 ]--------+--------------------------------------------------
 query               | INSERT INTO t3 VALUES (generate_series (?,?), ?);
 rows                | 10000
 total_time          | 52727.846
 shared_blks_dirtied | 10003
-[ RECORD 5 ]--------+--------------------------------------------------
 query               | INSERT INTO t4 VALUES (generate_series (?,?), ?);
 rows                | 10000
 total_time          | 56063.697
 shared_blks_dirtied | 10003

The times of session 1 (28407.435) and session 2 (31343.458) are close to each other, as they are both in the same resource group with dirty_rate_limit set to 6144. These times differe from the times of session 3 (52727.846) and session 4 (56063.697), which are in the resource group with dirty_rate_limit set to 3072. The latter group has a slower dirty rate limit, so the expected processing time is longer, as is the case for sessions 3 and 4.

The actual dirty rate for session 1 is calculated as follows:

  • The number of blocks dirtied per millisecond (ms) is 10003 blocks / 28407.435 ms, which yields 0.35212612 blocks per millisecond.
  • Multiply the result by 1000 to give the number of shared blocks dirtied per second (1 second = 1000 ms), which yields 352.12612 blocks per second.
  • Multiply the result by 8.192 to give the number of kilobytes dirtied per second (1 block = 8.192 kilobytes), which yields approximately 2885 kilobytes per second.

The actual dirty rate for session 2 is calculated as follows:

  • The number of blocks dirtied per millisecond (ms) is 10003 blocks / 31343.458 ms, which yields 0.31914156 blocks per millisecond.
  • Multiply the result by 1000 to give the number of shared blocks dirtied per second (1 second = 1000 ms), which yields 319.14156 blocks per second.
  • Multiply the result by 8.192 to give the number of kilobytes dirtied per second (1 block = 8.192 kilobytes), which yields approximately 2614 kilobytes per second.

The combined dirty rate from session 1 (2885 kilobytes per second) and from session 2 (2614 kilobytes per second) yields 5499 kilobytes per second, which is near the set dirty rate limit of the resource group (6144 kilobytes per seconds).

The actual dirty rate for session 3 is calculated as follows:

  • The number of blocks dirtied per millisecond (ms) is 10003 blocks / 52727.846 ms, which yields 0.18971001 blocks per millisecond.
  • Multiply the result by 1000 to give the number of shared blocks dirtied per second (1 second = 1000 ms), which yields 189.71001 blocks per second.
  • Multiply the result by 8.192 to give the number of kilobytes dirtied per second (1 block = 8.192 kilobytes), which yields approximately 1554 kilobytes per second.

The actual dirty rate for session 4 is calculated as follows:

  • The number of blocks dirtied per millisecond (ms) is 10003 blocks / 56063.697 ms, which yields 0.17842205 blocks per millisecond.
  • Multiply the result by 1000 to give the number of shared blocks dirtied per second (1 second = 1000 ms), which yields 178.42205 blocks per second.
  • Multiply the result by 8.192 to give the number of kilobytes dirtied per second (1 block = 8.192 kilobytes), which yields approximately 1462 kilobytes per second.

The combined dirty rate from session 3 (1554 kilobytes per second) and from session 4 (1462 kilobytes per second) yields 3016 kilobytes per second, which is near the set dirty rate limit of the resource group (3072 kilobytes per seconds).

This example shows how EDB Resource Manager keeps the aggregate dirty rate of the active processes in its groups close to the dirty rate limit set for each group.

← Prev

CPU usage throttling

↑ Up

Throttling CPU and I/O at the process level

Next →

Loading bulk data