GPU 负载下 CGEventPost 的性能较弱

Question

我们偶然发现了 Quartz Events 的性能问题，更具体地说是 CGEventPost：在 GPU 负载过重时，CGEventPost 可能会阻塞。我们创建了a small benchmark application to demonstrate the issue。这个应用程序只是一个创建、发布和发布事件的循环。

您可以在下面看到运行应用程序的结果。第一次运行是在空闲系统上。第二次运行是使用 FurMark（GPU 压力测试），表盘尽可能地向上转动。

• Inner 是内部循环所花费的时间，基本上只是使用 Quartz Events 创建、发布和发布事件。
• Outer 是我们的程序等待唤醒（睡眠）的时间。应该接近我们睡觉的时间，但如果系统处于压力之下，它可能会延迟。
• 发布是事件发布所需的时间。

 

18:58:01.683 EventPerformance[4946:707] Measurements: (outer should be close to 10)
18:58:01.684 EventPerformance[4946:707] inner (ms): 0.04, outer (ms): 11.02, CGEventPost (ms): 0.03
18:58:01.684 EventPerformance[4946:707] inner (ms): 0.04, outer (ms): 11.02, CGEventPost (ms): 0.03
18:58:01.685 EventPerformance[4946:707] inner (ms): 0.07, outer (ms): 10.26, CGEventPost (ms): 0.03
18:58:01.685 EventPerformance[4946:707] inner (ms): 0.06, outer (ms): 10.85, CGEventPost (ms): 0.05
18:58:01.686 EventPerformance[4946:707] inner (ms): 0.07, outer (ms): 10.41, CGEventPost (ms): 0.04
18:58:01.686 EventPerformance[4946:707] inner (ms): 0.04, outer (ms): 10.39, CGEventPost (ms): 0.03
18:58:01.686 EventPerformance[4946:707] inner (ms): 0.05, outer (ms): 11.02, CGEventPost (ms): 0.03
18:58:01.687 EventPerformance[4946:707] inner (ms): 0.03, outer (ms): 10.67, CGEventPost (ms): 0.03
18:58:01.687 EventPerformance[4946:707] inner (ms): 0.08, outer (ms): 10.09, CGEventPost (ms): 0.05
18:58:01.688 EventPerformance[4946:707] Averages: (outer should be close to 10)
18:58:01.688 EventPerformance[4946:707] avg inner (ms): 0.05, avg outer (ms): 10.64, avg post (ms): 0.03

在这里我们可以看到发布事件平均需要大约 0.03 毫秒。此外，线程似乎在 0.5ms 左右被唤醒为时已晚。 CGEventPost 中没有峰值。

19:02:02.150 EventPerformance[5241:707] Measurements: (outer should be close to 10)
19:02:02.151 EventPerformance[5241:707] inner (ms): 0.03, outer (ms): 10.23, CGEventPost (ms): 0.02
19:02:02.151 EventPerformance[5241:707] inner (ms): 0.02, outer (ms): 10.54, CGEventPost (ms): 0.02
19:02:02.151 EventPerformance[5241:707] inner (ms): 0.02, outer (ms): 11.01, CGEventPost (ms): 0.01
19:02:02.152 EventPerformance[5241:707] inner (ms): 0.02, outer (ms): 10.74, CGEventPost (ms): 0.01
19:02:02.152 EventPerformance[5241:707] inner (ms): 0.02, outer (ms): 10.20, CGEventPost (ms): 0.01
19:02:02.152 EventPerformance[5241:707] inner (ms): 10.35, outer (ms): 11.01, CGEventPost (ms): 10.35
19:02:02.152 EventPerformance[5241:707] inner (ms): 0.03, outer (ms): 10.02, CGEventPost (ms): 0.02
19:02:02.153 EventPerformance[5241:707] inner (ms): 58.90, outer (ms): 10.11, CGEventPost (ms): 58.90
19:02:02.153 EventPerformance[5241:707] inner (ms): 0.03, outer (ms): 10.12, CGEventPost (ms): 0.02
19:02:02.153 EventPerformance[5241:707] Averages: (outer should be close to 10)
19:02:02.371 EventPerformance[5241:707] avg inner (ms): 7.71, avg outer (ms): 10.44, avg post (ms): 7.71

当系统处于繁重的 GPU 负载下时，发布事件可能需要（峰值）毫秒而不是微秒。在极端 GPU 压力（有时似乎在等待 GPU 完成一些工作后再返回。我们的线程仍然正常调度，没有明显的延迟/尖峰（外部）。

感谢任何想法。

Answer 1

我猜你正在填满队列（底层 mach 端口）......

您可以使用 Instruments 中的“调度”或“系统调用”工具来确认这一点。 （创建一个新的空白文档，添加仪器，然后在File > Record Options... 下确保选中“延迟模式”。）这将显示应用程序中的所有线程活动（线程阻塞时、睡眠时、激活时以及为什么）。

我会首先尝试提高调用CGEventPost 的线程的线程优先级（参见man 3 PTHREAD_SCHEDPARAM）。如果您的线程在较低优先级的线程上被阻塞，内核应暂时提升阻塞线程的优先级以避免优先级倒置并帮助您的任务更早完成。

总体而言，我认为您将必须实现 2 线程解决方案，如下所示：

为您要发布的事件创建一个队列。从您的主线程（或事件发布线程）向此队列发布事件，然后向第二个线程（您创建的事件使用者线程）发出信号以遍历队列并使用CGEventPost 发布任何未完成的事件。

当CGEventPost 阻塞时，您的第二个事件发布线程将阻塞，但这不会阻塞任何其他线程。当CGEventPost 最终解除阻塞时，它将消耗您的事件消费者线程发布的任何未完成的事件，并且事件消费者线程可以恢复发布事件。

另一种可能性：你能合并事件吗？您可以将某些类型的事件（鼠标移动？）合并成更少的事件。有时您可能仍会遇到CGEventPost 的队列限制，我认为2 线程方法可能是您最好的选择。