[
https://issues.apache.org/jira/browse/KUDU-3557?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Song Jiacheng updated KUDU-3557:
--------------------------------
Description:
{color:#172b4d}Now, Lock contention is too high in tcmalloc while the cluster
in under high throughput. Making the total thread cache size larger could
highly decrease the lock contention. Even though the max size of one single
thread cache could not be larger than 4096KB, but tcmalloc will have much more
room to control the total size if {color}max_total_thread_cache_bytes is larger.
Here are some comparison between tservers which are set to different
max_total_thread_cache_bytes
ta2:512MB others:32MB(default)
x: time y: metric spinlock_contention_time
{color:#172b4d}!image-2024-02-27-15-58-05-449.png!{color}
{color:#172b4d}!image-2024-02-27-15-59-25-180.png!{color}
{color:#172b4d} {color}
was:
{color:#172b4d}Now, Lock contention is too high in tcmalloc while the cluster
in under high throughput. Making the total thread cache size larger could
highly decrease the lock contention. Even though the max size of one single
thread cache could not be larger than 4096KB, but tcmalloc will have much more
room to control the total size if {color}max_total_thread_cache_bytes is larger.
Here are some comparison between tservers which are set to different
max_total_thread_cache_bytes
ta2:512MB others:32MB(default)
{color:#172b4d}!image-2024-02-27-15-58-05-449.png!{color}
{color:#172b4d}!image-2024-02-27-15-59-25-180.png!{color}
{color:#172b4d} {color}
> Add ability to set maximum thread cache size of tcmalloc easily.
> ----------------------------------------------------------------
>
> Key: KUDU-3557
> URL: https://issues.apache.org/jira/browse/KUDU-3557
> Project: Kudu
> Issue Type: Improvement
> Reporter: Song Jiacheng
> Priority: Major
> Attachments: image-2024-02-27-15-58-05-449.png,
> image-2024-02-27-15-58-14-096.png, image-2024-02-27-15-59-25-180.png
>
>
> {color:#172b4d}Now, Lock contention is too high in tcmalloc while the cluster
> in under high throughput. Making the total thread cache size larger could
> highly decrease the lock contention. Even though the max size of one single
> thread cache could not be larger than 4096KB, but tcmalloc will have much
> more room to control the total size if {color}max_total_thread_cache_bytes is
> larger.
> Here are some comparison between tservers which are set to different
> max_total_thread_cache_bytes
> ta2:512MB others:32MB(default)
> x: time y: metric spinlock_contention_time
> {color:#172b4d}!image-2024-02-27-15-58-05-449.png!{color}
> {color:#172b4d}!image-2024-02-27-15-59-25-180.png!{color}
> {color:#172b4d} {color}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
