[DocDB] Gate new memory default recommendations on an ordinary gflag #22161
Assignees
Labels
2024.1.0_blocker
area/docdb
YugabyteDB core features
jira-originated
kind/new-feature
This is a request for a completely new feature
priority/high
High Priority
Comments
yugabyte-ci
added
area/docdb
yugabyte-ci
added
priority/high
|
Currently these are turned on for testing; we decided for 2024.1 to turn them on by default but rather require customers to opt in by setting a gflag. |
mdbridge
added a commit
that referenced
this issue
Apr 26, 2024
…y gflag
Summary:
Here we introduce a new gflag:
```
~/code/yugabyte-db/src/yb/tserver/server_main_util.cc:36:
DEFINE_NON_RUNTIME_bool(
use_memory_defaults_optimized_for_ysql, false,
"If true, the recommended defaults for the memory usage settings take into account the amount "
"of RAM and cores available and are optimized for using YSQL. "
"If false, the recommended defaults will be the old defaults, which are more suitable "
"for YCQL but do not take in account the amount of RAM and cores available.");
```
to gate the new memory recommendations.
More precisely, we still have a concept of recommended defaults and individual memory settings can choose whether or not to use the recommended defaults. For example:
```
~/code/yugabyte-db/src/yb/tserver/tablet_memory_manager.cc:59:
// NOTE: The default here is for tools and tests; the actual defaults
// for the TServer and master processes are set in server_main_util.cc.
DEFINE_NON_RUNTIME_int32(tablet_overhead_size_percentage, 0,
"Percentage of total available memory to use for tablet-related overheads. A value of 0 means "
"no limit. Must be between 0 and 100 inclusive. Exception: "
BOOST_PP_STRINGIZE(USE_RECOMMENDED_MEMORY_VALUE) " specifies to instead use a "
"recommended value determined in part by the amount of RAM available.");
```
If this gflag is set to USE_RECOMMENDED_MEMORY_VALUE by the customer or left unset (the gflag default is set to that value for the TServer and master processes elsewhere), then we will use the recommended default for it.
If the new gflag is set to true then we use the new recommendations we have derived using our models. If it is set to false, then we use the same values we would've used prior to version 2024.1 if the customer had not provided a value for this gflag.
Thus, the customer gets essentially the same behavior as before 2024.1 if they do not set the new gflag.
Jira: DB-11090
Test Plan:
Started cluster with and without flag on and look at logs:
```
~/code/yugabyte-db/bin/yb-ctl start
I0425 10:52:25.180203 5935 server_main_util.cc:90] Setting flag db_block_cache_size_percentage to recommended value -3
I0425 10:52:25.180212 5935 server_main_util.cc:92] Flag default_memory_limit_to_ram_ratio has value 0.65 (recommended value is 0.85)
I0425 10:52:25.180217 5935 server_main_util.cc:94] Setting flag tablet_overhead_size_percentage to recommended value 0
```
```
bin/yb-ctl start --tserver_flags "use_memory_defaults_optimized_for_ysql=true" \
--master_flag "use_memory_defaults_optimized_for_ysql=true"
I0425 10:57:00.198920 6412 server_main_util.cc:101] Total available RAM is 31.183224 GiB
I0425 10:57:00.198928 6412 server_main_util.cc:90] Setting flag db_block_cache_size_percentage to recommended value 32
I0425 10:57:00.198938 6412 server_main_util.cc:92] Flag default_memory_limit_to_ram_ratio has value 0.65 (recommended value is 0.6)
I0425 10:57:00.198943 6412 server_main_util.cc:94] Setting flag tablet_overhead_size_percentage to recommended value 10
```
(yb-ctl overrides default_memory_limit_to_ram_ratio.)
Reviewers: zdrudi
Reviewed By: zdrudi
Subscribers: ybase
Differential Revision: https://phorge.dev.yugabyte.com/D34547
mdbridge
added a commit
that referenced
this issue
Apr 27, 2024
…ons on an ordinary gflag Summary: Original commit: b64cac3 / D34547 Clean cherry pick, no conflicts. Here we introduce a new gflag: ``` ~/code/yugabyte-db/src/yb/tserver/server_main_util.cc:36: DEFINE_NON_RUNTIME_bool( use_memory_defaults_optimized_for_ysql, false, "If true, the recommended defaults for the memory usage settings take into account the amount " "of RAM and cores available and are optimized for using YSQL. " "If false, the recommended defaults will be the old defaults, which are more suitable " "for YCQL but do not take in account the amount of RAM and cores available."); ``` to gate the new memory recommendations. More precisely, we still have a concept of recommended defaults and individual memory settings can choose whether or not to use the recommended defaults. For example: ``` ~/code/yugabyte-db/src/yb/tserver/tablet_memory_manager.cc:59: // NOTE: The default here is for tools and tests; the actual defaults // for the TServer and master processes are set in server_main_util.cc. DEFINE_NON_RUNTIME_int32(tablet_overhead_size_percentage, 0, "Percentage of total available memory to use for tablet-related overheads. A value of 0 means " "no limit. Must be between 0 and 100 inclusive. Exception: " BOOST_PP_STRINGIZE(USE_RECOMMENDED_MEMORY_VALUE) " specifies to instead use a " "recommended value determined in part by the amount of RAM available."); ``` If this gflag is set to USE_RECOMMENDED_MEMORY_VALUE by the customer or left unset (the gflag default is set to that value for the TServer and master processes elsewhere), then we will use the recommended default for it. If the new gflag is set to true then we use the new recommendations we have derived using our models. If it is set to false, then we use the same values we would've used prior to version 2024.1 if the customer had not provided a value for this gflag. Thus, the customer gets essentially the same behavior as before 2024.1 if they do not set the new gflag. Jira: DB-11090 Test Plan: Started cluster with and without flag on and look at logs: ``` ~/code/yugabyte-db/bin/yb-ctl start I0425 10:52:25.180203 5935 server_main_util.cc:90] Setting flag db_block_cache_size_percentage to recommended value -3 I0425 10:52:25.180212 5935 server_main_util.cc:92] Flag default_memory_limit_to_ram_ratio has value 0.65 (recommended value is 0.85) I0425 10:52:25.180217 5935 server_main_util.cc:94] Setting flag tablet_overhead_size_percentage to recommended value 0 ``` ``` bin/yb-ctl start --tserver_flags "use_memory_defaults_optimized_for_ysql=true" \ --master_flag "use_memory_defaults_optimized_for_ysql=true" I0425 10:57:00.198920 6412 server_main_util.cc:101] Total available RAM is 31.183224 GiB I0425 10:57:00.198928 6412 server_main_util.cc:90] Setting flag db_block_cache_size_percentage to recommended value 32 I0425 10:57:00.198938 6412 server_main_util.cc:92] Flag default_memory_limit_to_ram_ratio has value 0.65 (recommended value is 0.6) I0425 10:57:00.198943 6412 server_main_util.cc:94] Setting flag tablet_overhead_size_percentage to recommended value 10 ``` (yb-ctl overrides default_memory_limit_to_ram_ratio.) Reviewers: zdrudi, slingam Reviewed By: slingam Subscribers: ybase Tags: #jenkins-ready Differential Revision: https://phorge.dev.yugabyte.com/D34565
svarnau
pushed a commit
that referenced
this issue
May 25, 2024
…y gflag
Summary:
Here we introduce a new gflag:
```
~/code/yugabyte-db/src/yb/tserver/server_main_util.cc:36:
DEFINE_NON_RUNTIME_bool(
use_memory_defaults_optimized_for_ysql, false,
"If true, the recommended defaults for the memory usage settings take into account the amount "
"of RAM and cores available and are optimized for using YSQL. "
"If false, the recommended defaults will be the old defaults, which are more suitable "
"for YCQL but do not take in account the amount of RAM and cores available.");
```
to gate the new memory recommendations.
More precisely, we still have a concept of recommended defaults and individual memory settings can choose whether or not to use the recommended defaults. For example:
```
~/code/yugabyte-db/src/yb/tserver/tablet_memory_manager.cc:59:
// NOTE: The default here is for tools and tests; the actual defaults
// for the TServer and master processes are set in server_main_util.cc.
DEFINE_NON_RUNTIME_int32(tablet_overhead_size_percentage, 0,
"Percentage of total available memory to use for tablet-related overheads. A value of 0 means "
"no limit. Must be between 0 and 100 inclusive. Exception: "
BOOST_PP_STRINGIZE(USE_RECOMMENDED_MEMORY_VALUE) " specifies to instead use a "
"recommended value determined in part by the amount of RAM available.");
```
If this gflag is set to USE_RECOMMENDED_MEMORY_VALUE by the customer or left unset (the gflag default is set to that value for the TServer and master processes elsewhere), then we will use the recommended default for it.
If the new gflag is set to true then we use the new recommendations we have derived using our models. If it is set to false, then we use the same values we would've used prior to version 2024.1 if the customer had not provided a value for this gflag.
Thus, the customer gets essentially the same behavior as before 2024.1 if they do not set the new gflag.
Jira: DB-11090
Test Plan:
Started cluster with and without flag on and look at logs:
```
~/code/yugabyte-db/bin/yb-ctl start
I0425 10:52:25.180203 5935 server_main_util.cc:90] Setting flag db_block_cache_size_percentage to recommended value -3
I0425 10:52:25.180212 5935 server_main_util.cc:92] Flag default_memory_limit_to_ram_ratio has value 0.65 (recommended value is 0.85)
I0425 10:52:25.180217 5935 server_main_util.cc:94] Setting flag tablet_overhead_size_percentage to recommended value 0
```
```
bin/yb-ctl start --tserver_flags "use_memory_defaults_optimized_for_ysql=true" \
--master_flag "use_memory_defaults_optimized_for_ysql=true"
I0425 10:57:00.198920 6412 server_main_util.cc:101] Total available RAM is 31.183224 GiB
I0425 10:57:00.198928 6412 server_main_util.cc:90] Setting flag db_block_cache_size_percentage to recommended value 32
I0425 10:57:00.198938 6412 server_main_util.cc:92] Flag default_memory_limit_to_ram_ratio has value 0.65 (recommended value is 0.6)
I0425 10:57:00.198943 6412 server_main_util.cc:94] Setting flag tablet_overhead_size_percentage to recommended value 10
```
(yb-ctl overrides default_memory_limit_to_ram_ratio.)
Reviewers: zdrudi
Reviewed By: zdrudi
Subscribers: ybase
Differential Revision: https://phorge.dev.yugabyte.com/D34547
ZhenYongFan
pushed a commit
to ZhenYongFan/yugabyte-db
that referenced
this issue
Jun 15, 2024
…mmendations on an ordinary gflag Summary: Original commit: b64cac3 / D34547 Clean cherry pick, no conflicts. Here we introduce a new gflag: ``` ~/code/yugabyte-db/src/yb/tserver/server_main_util.cc:36: DEFINE_NON_RUNTIME_bool( use_memory_defaults_optimized_for_ysql, false, "If true, the recommended defaults for the memory usage settings take into account the amount " "of RAM and cores available and are optimized for using YSQL. " "If false, the recommended defaults will be the old defaults, which are more suitable " "for YCQL but do not take in account the amount of RAM and cores available."); ``` to gate the new memory recommendations. More precisely, we still have a concept of recommended defaults and individual memory settings can choose whether or not to use the recommended defaults. For example: ``` ~/code/yugabyte-db/src/yb/tserver/tablet_memory_manager.cc:59: // NOTE: The default here is for tools and tests; the actual defaults // for the TServer and master processes are set in server_main_util.cc. DEFINE_NON_RUNTIME_int32(tablet_overhead_size_percentage, 0, "Percentage of total available memory to use for tablet-related overheads. A value of 0 means " "no limit. Must be between 0 and 100 inclusive. Exception: " BOOST_PP_STRINGIZE(USE_RECOMMENDED_MEMORY_VALUE) " specifies to instead use a " "recommended value determined in part by the amount of RAM available."); ``` If this gflag is set to USE_RECOMMENDED_MEMORY_VALUE by the customer or left unset (the gflag default is set to that value for the TServer and master processes elsewhere), then we will use the recommended default for it. If the new gflag is set to true then we use the new recommendations we have derived using our models. If it is set to false, then we use the same values we would've used prior to version 2024.1 if the customer had not provided a value for this gflag. Thus, the customer gets essentially the same behavior as before 2024.1 if they do not set the new gflag. Jira: DB-11090 Test Plan: Started cluster with and without flag on and look at logs: ``` ~/code/yugabyte-db/bin/yb-ctl start I0425 10:52:25.180203 5935 server_main_util.cc:90] Setting flag db_block_cache_size_percentage to recommended value -3 I0425 10:52:25.180212 5935 server_main_util.cc:92] Flag default_memory_limit_to_ram_ratio has value 0.65 (recommended value is 0.85) I0425 10:52:25.180217 5935 server_main_util.cc:94] Setting flag tablet_overhead_size_percentage to recommended value 0 ``` ``` bin/yb-ctl start --tserver_flags "use_memory_defaults_optimized_for_ysql=true" \ --master_flag "use_memory_defaults_optimized_for_ysql=true" I0425 10:57:00.198920 6412 server_main_util.cc:101] Total available RAM is 31.183224 GiB I0425 10:57:00.198928 6412 server_main_util.cc:90] Setting flag db_block_cache_size_percentage to recommended value 32 I0425 10:57:00.198938 6412 server_main_util.cc:92] Flag default_memory_limit_to_ram_ratio has value 0.65 (recommended value is 0.6) I0425 10:57:00.198943 6412 server_main_util.cc:94] Setting flag tablet_overhead_size_percentage to recommended value 10 ``` (yb-ctl overrides default_memory_limit_to_ram_ratio.) Reviewers: zdrudi, slingam Reviewed By: slingam Subscribers: ybase Tags: #jenkins-ready Differential Revision: https://phorge.dev.yugabyte.com/D34565
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Jira Link: DB-11090
The text was updated successfully, but these errors were encountered: