在性能问题之前有多少并发 setTimeouts?

IT技术 javascript node.js performance settimeout
2021-02-24 11:36:29

我有一个 node.js 应用程序,在任何给定时间运行 10k-100k 并发 setTimeouts。(它们都是 5 分钟的持续时间。)回调非常简单,只是 redis 中的一个 HDECRBY。我还没有遇到任何性能问题,即使在 t2.micro 实例上也是如此。

我知道我会遇到问题,如果回调函数不能得到尽可能快,因为我设置一个定时器,(显然)执行,但是否有问题,具有高数一个定时器,的本身例如,如果我将其扩展到 100 万并发,我会遇到 RAM 瓶颈吗?千万?

1个回答

对于这些类型的问题,查看 node.js 如何处理源代码中的计时器通常很有用

你会发现 node.js 保存了一个或多个它自己的内部计时器对象的链表,并且所有设置为同时发生的计时器共享一个 libuv 计时器。这意味着所有设置为在相当特定的时间窗口中发生的无数计时器将不可避免地共享许多触发时间,因此将共享计时器列表,从而将共享许多系统计时器对象。

这使得拥有数以百万计的计时器对象不再是一个问题。现在,每个计时器对象仍然需要一些内存,并不是计时器实现中的每个操作都是恒定时间,尽管您可以在下面的评论中看到,他们试图使尽可能多的它们成为恒定时间以允许大量计时器性能仍然不错。

如果您在计时器触发时不需要绝对精度,您可以通过仅将计时器安排在特定的时间边界(例如 100 毫秒的偶数),从而使计时器更频繁地合并和共享计时器对象。这将在相同的触发时间内安排更多的计时器,并允许 node.js 将更多计时器放入同一个列表中,所有计时器都共享一个系统计时器。我不知道这对您的计时器是否可行,或者是否需要,但在研究 node.js 的工作原理时,它会提高效率。node.js 内部的计时器列表会更少,而 libuv 中的系统计时器也会更少。


以下是 node.js 代码中关于计时器的一些解释性注释,解释了设计的更多方面:

// HOW and WHY the timers implementation works the way it does.
//
// Timers are crucial to Node.js. Internally, any TCP I/O connection creates a
// timer so that we can time out of connections. Additionally, many user
// user libraries and applications also use timers. As such there may be a
// significantly large amount of timeouts scheduled at any given time.
// Therefore, it is very important that the timers implementation is performant
// and efficient.
//
// Note: It is suggested you first read though the lib/internal/linkedlist.js
// linked list implementation, since timers depend on it extensively. It can be
// somewhat counter-intuitive at first, as it is not actually a class. Instead,
// it is a set of helpers that operate on an existing object.
//
// In order to be as performant as possible, the architecture and data
// structures are designed so that they are optimized to handle the following
// use cases as efficiently as possible:

// - Adding a new timer. (insert)
// - Removing an existing timer. (remove)
// - Handling a timer timing out. (timeout)
//
// Whenever possible, the implementation tries to make the complexity of these
// operations as close to constant-time as possible.
// (So that performance is not impacted by the number of scheduled timers.)
//
// Object maps are kept which contain linked lists keyed by their duration in
// milliseconds.
// The linked lists within also have some meta-properties, one of which is a
// TimerWrap C++ handle, which makes the call after the duration to process the
// list it is attached to.
//
//
// ╔════ > Object Map
// ║
// ╠══
// ║ refedLists: { '40': { }, '320': { etc } } (keys of millisecond duration)
// ╚══          ┌─────────┘
//              │
// ╔══          │
// ║ TimersList { _idleNext: { }, _idlePrev: (self), _timer: (TimerWrap) }
// ║         ┌────────────────┘
// ║    ╔══  │                              ^
// ║    ║    { _idleNext: { },  _idlePrev: { }, _onTimeout: (callback) }
// ║    ║      ┌───────────┘
// ║    ║      │                                  ^
// ║    ║      { _idleNext: { etc },  _idlePrev: { }, _onTimeout: (callback) }
// ╠══  ╠══
// ║    ║
// ║    ╚════ >  Actual JavaScript timeouts
// ║
// ╚════ > Linked List
//
//
// With this, virtually constant-time insertion (append), removal, and timeout
// is possible in the JavaScript layer. Any one list of timers is able to be
// sorted by just appending to it because all timers within share the same
// duration. Therefore, any timer added later will always have been scheduled to
// timeout later, thus only needing to be appended.
// Removal from an object-property linked list is also virtually constant-time
// as can be seen in the lib/internal/linkedlist.js implementation.
// Timeouts only need to process any timers due to currently timeout, which will
// always be at the beginning of the list for reasons stated above. Any timers
// after the first one encountered that does not yet need to timeout will also
// always be due to timeout at a later time.
//
// Less-than constant time operations are thus contained in two places:
// TimerWrap's backing libuv timers implementation (a performant heap-based
// queue), and the object map lookup of a specific list by the duration of
// timers within (or creation of a new list).
// However, these operations combined have shown to be trivial in comparison to
// other alternative timers architectures.


// Object maps containing linked lists of timers, keyed and sorted by their
// duration in milliseconds.
//
// The difference between these two objects is that the former contains timers
// that will keep the process open if they are the only thing left, while the
// latter will not.
另一个我没有想过的问题是存储所有的回调函数。每个 setTimeout 都是在回调期间设置的,所以我必须使用闭包...所以这意味着内存中漂浮着数以千计的匿名函数,同一个函数的所有副本只是具有不同的变量。也许不是使用闭包(不带参数)“setTimeout(closure(x),30000)”,我应该为每个 setTimeout 使用相同的函数,并使用闭包将参数传递给它,如“setTimeout(func,30000,closure (x))" 以便成千上万的匿名函数在内存中更小。
2021-04-25 11:36:29
我提到我所有的计时器的持续时间都是 5 分钟(30000 毫秒),所以这应该意味着我的所有计时器都包含在一个链接列表中并共享一个 libuv 计时器。伟大的。
2021-05-05 11:36:29
@barron 它们是按计时器类型组织的,而不是按持续时间组织的。
2021-05-09 11:36:29
@barron - 这取决于您的所有计时器是否同时设置为 30000 毫秒。计时器是按触发时间(计时器设置为触发的时钟时间)收集的,而不是从设置计时器的那一刻起的持续时间。因此,如果您的所有计时器都设置在同一时间,则它们将被共享,但如果它们在不同的时间设置,那么它们将在不同的时间触发,并且不会被组织在一起。
2021-05-13 11:36:29
@zerkms“保留对象映射,其中包含以毫秒为单位的持续时间键控的链接列表。” 我确定我在这里误解了一些东西。
2021-05-15 11:36:29