push-gently

A push-stream through that applies backpressure when CPU is too busy

npm install --save push-gently

Similar to pull-drain-gently, but for push-streams.

Usage

Default use

const gently = require('push-gently');

myPushStream
  .pipe(gently())
  .pipe({
    paused: false,
    write: function (x) {
      console.log(x);
    },
  });

Tweaking the parameters

This push-stream through listens to the source, but once in a while it will check the CPU usage, and if it has gone above the limit known as the ceiling, it pauses itself for wait milliseconds, and then resumes the source, in such a way that it should stay close to the limit. In other words, two parameters control the draining:

• ceiling: the maximum CPU usage you want the Node.js process to consume, approximately, in percentages (100 is 100%, not 1)
• wait: the waiting period, in milliseconds, to pause draining in order to allow other tasks to use the CPU

The default CPU ceiling is 88% and the default waiting period for each pause is 144ms (roughly 9 frames if you have the UI running at 60fps). This means that the longest pause during which no drains will occur is 144ms.

There is also a third less common parameter, which by default is turned off:

• maxPause: a limit in milliseconds for how long to pause. For instance, when maxPause = 5000, the draining will remain paused for 5 seconds maximum, after that it will d resume draining regardless of the current CPU usage. The default is Infinity.

To configure your own parameters, pass an opts object as the argument to gently:

const gently = require('push-gently');

myPushStream
  .pipe(gently({ ceiling: 90, wait: 60 })) // <--- opts
  .pipe({
    paused: false,
    write: function (x) {
      console.log(x);
    },
  });

To configure these parameters, consider that:

• The greater the ceiling is, the closer the pipeline behaves to stock push.drain, i.e. the less gently it drains with regards to CPU load
• The smaller the ceiling is, the more time it will take to consume the source, i.e. the slower your application will finish its CPU tasks
• The greater the wait, the more the actual CPU usage fluctuates below and above the ceiling, i.e. the more bumpy the ride is for CPU usage and workload throughput
• The smaller the wait, the more the actual CPU usage accurately meets the ceiling, but also the more overhead there is with many short-lived timers for those pauses

The total time for drainage is also important. push.drain without gently is the fastest, having the shortest total time. gently with a small wait might give a total drainage time of approx. 2.5x that of push.drain.

The defaults ceiling=88, wait=144 are a sweet spot, and it can achieve a total drainage time of approx 1.4x that of push.drain. You can run benchmarks yourself by running cd perf && ./run-all.sh in this repository.

The table below shows how the JS CPU flamechart looks like when running the benchmark in perf on Ubuntu 18.04.3 x86_64, Intel® Core™ i7-7500U CPU @ 2.70GHz × 4, 15,4 GiB RAM, for different values of the ceiling parameter. "Unlimited" means push.drain without gently:

| CPU ceiling | CPU Profiler in Chrome | |-----------|------------------------| | Unlimited | | | 90% | | | 80% | | | 70% | | | 60% | | | 50% | | | 40% | |

License

MIT