tencent cloud

Feedback

Dynamic Thread Pool

Last updated: 2024-02-18 11:24:04

    Overview

    The thread pool (Thread_pool) uses a certain number of worker threads to process connection requests, which is typically used in scenarios with online transaction processing (OLTP) workloads. However, when many requests are slow queries, worker threads will be blocked by high-latency operations and fail to quickly respond to new requests. As a result, the system throughput of the thread pool mode is lower than that of the traditional one-thread-per-connection (Per_thread) mode.
    The Per_thread and Thread_pool modes have their advantages and disadvantages, so the system needs to flexibly switch between them based on business types. Unfortunately, the mode switch must be completed by restarting the server (during peak hours in most cases), adversely affecting the business.
    To allow users to flexibly switch between Per_thread and Thread_pool, TencentDB for MySQL has introduced the optimization of thread pool dynamic switch, that is, to enable or disable the thread pool without restarting the database service.

    Supported Versions

    Kernel version: MySQL 8.0 20201230 and above.
    Kernel version: MySQL 5.7 20201230 and above.

    Use Cases

    This feature is suitable for the business which is sensitive to performance and needs to flexibly change the database working mode based on the business type.

    Performance Impact

    Switching from the thread pool mode to the one-thread-per-connection mode won't block queries or affect database performance.
    Switching from the one-thread-per-connection mode to the thread pool mode under extremely high QPS and persistent high pressure may block requests because the thread pool is disabled before the switch.
    Solution 1: you can increase the thread_pool_oversubscribe parameter and decrease the thread_pool_stall_limit parameter to quickly enable the thread pool. After the blocked SQL queries are processed, you can restore the parameters to their original values as needed.
    Solution 2: if SQL queries start to be blocked, you can suspend or reduce service traffic for a few seconds, wait for thread pool enablement to complete, and then resume the continuous high-pressure service traffic.

    Instructions

    You can use the thread_handling_switch_mode parameter to control whether to dynamically change the thread working mode. Parameter values are described as follows:
    Valid Value
    Description
    disabled
    The mode cannot be changed dynamically.
    stable
    The mode can only be changed for new connections.
    fast
    (Default value) The mode can be changed for new connections and new requests.
    sharp
    Active connections will be killed in order to force the user to reconnect so that the mode can be changed quickly.
    The show threadpool status command displays the following new status:
    connections_moved_from_per_thread: the number of connections switched from Per_thread to Thread_pool.
    connections_moved_to_per_thread: the number of connections switched from Thread_pool to Per_thread.
    events_consumed: the total number of events consumed by the worker thread group in each thread pool. After the thread working mode is switched from Thread_pool to Per_thread, the total number of events won't increase any more.
    average_wait_usecs_in_queue: the average time each event waits in the queue.
    The show full processlist command displays the following new status:
    Moved_to_per_thread: the number of times that the connection is switched to Per_thread.
    Moved_to_thread_pool: the number of times that the connection is switched to Thread_pool.

    Parameter Status Description

    Thread pool parameters are described as follows:
    Parameter
    Effective Immediately
    Type
    Default Value
    Valid Values/Value Range
    Description
    thread_pool_idle_timeout
    Yes
    uint
    60
    [1, UINT_MAX]
    If there are no network event to handle, the worker thread will wait for thread_pool_idle_timeout seconds before being killed
    thread_pool_oversubscribe
    Yes
    uint
    High Stability Parameter Template: 10
    High Performance Parameter Template: 16
    [3,32]
    The maximum number of worker threads allowed in a worker thread group
    thread_pool_size
    Yes
    uint
    The number of CPUs on the current machine
    [1,1000]
    The number of thread groups
    thread_pool_stall_limit
    Yes
    uint
    500
    [10, UINT_MAX]
    The timer thread checks all thread groups every thread_pool_stall_limit milliseconds.
    If a thread group has no listener, neither high nor low priority queue is empty, and no new I/O network event occurs, the thread group is considered stalled. The timer thread will wake up a sleeping thread or create a new thread to relieve the thread group's pressure.
    thread_pool_max_threads
    Yes
    uint
    100000
    [1,100000]
    The total number of worker threads in the thread pool
    thread_pool_high_prio_mode
    Yes, session
    enum
    transactions
    transactions\\statement\\none
    The high priority queue supports three modes:
    transactions: only when there is one statement that already starts a transaction and the ticket (thread_pool_high_prio_tickets) held by the connection of this statement is not zero, the connection can enter the high priority queue. After a connection has entered the priority queue thread_pool_high_prio_tickets times, it will be placed into the ordinary queue.
    statement: all connections enter the high priority queue.
    none: all connections enter the low priority queue.
    thread_pool_high_prio_tickets
    Yes, session
    uint
    UINT_MAX
    [0, UINT_MAX]
    The value of the ticket assigned to each connection in transactions mode
    threadpool_workaround_epoll_bug
    Yes
    bool
    false
    true/false
    Whether to bypass the epoll bug of Linux 2.x (which was fixed in Linux 3)
    The show threadpool status command displays the following status:
    Status
    Description
    groupid
    Thread group ID
    connection_count
    The number of user connections in the thread group
    thread_count
    The number of worker threads in the thread group
    havelistener
    Whether the thread group has a listener
    active_thread_count
    The number of active worker threads in the thread group
    waiting_thread_count
    The number of worker threads calling wait_begin in the thread group
    waiting_threads_size
    The number of sleeping worker threads waiting to be woken up in the thread group when there is no network event to handle (such worker threads will wait for thread_pool_idle_timeout seconds before being automatically killed)
    queue_size
    The length of the ordinary queue of the thread group
    high_prio_queue_size
    The length of the high priority queue of the thread group
    get_high_prio_queue_num
    The total number of times that events in the thread group are removed from the high priority queue
    get_normal_queue_num
    The total number of times that events in the thread group are removed from the ordinary queue
    create_thread_num
    The total number of worker threads created in the thread group
    wake_thread_num
    The total number of worker threads in the thread group awakened from the waiting_threads queue
    oversubscribed_num
    The number of times that worker threads are ready to go to sleep because the thread group is oversubscribed
    mysql_cond_timedwait_num
    The total number of times that worker threads in the thread group enter the waiting_threads queue
    check_stall_nolistener
    The total number of times that no listener is detected in the thread group in the stall check performed by the timer thread
    check_stall_stall
    The total number of times that the thread group is considered stalled in the stall check performed by the timer thread
    max_req_latency_us
    The maximum time in milliseconds for a user connection to wait in the queue in the thread group
    conns_timeout_killed
    The total number of times that user connections in the thread group are killed because there has been no new message on the client for the threshold period (net_wait_timeout)
    connections_moved_in
    The total number of connections migrated from other thread groups to this thread group
    connections_moved_out
    The total number of connections migrated from this thread group to other thread groups
    connections_moved_from_per_thread
    The total number of connections switched from the one-thread-per-connection mode to this thread group
    connections_moved_to_per_thread
    The total number of connections switched from this thread group to the one-thread-per-connection mode
    events_consumed
    The total number of events processed by the thread group
    average_wait_usecs_in_queue
    The average waiting time of all events in the queue in the thread group
    
    Contact Us

    Contact our sales team or business advisors to help your business.

    Technical Support

    Open a ticket if you're looking for further assistance. Our Ticket is 7x24 avaliable.

    7x24 Phone Support