According to https://nodejs.org/api/cluster.html#cluster_cluster, one should run the same number of Node.js processes in parallel as the number of cores on the machine. The supposed reasoning behind this is that Node.js is single threaded.
However, is this really true? Sure the JavaScript code and the event loop run on one thread but Node also has a worker thread pool. The default number of thread in this pool is 4. So why does it make sense to run one Node process per core?
This article has an extension review on the threading mechanism of node.js, worth a read.
In short, the main point is in plain node.js only a few function calls uses thread pool (DNS and FS calls). Your call mostly runs on the event loop only. So for example if you wrote a web app that each request takes 100ms synchronously, you are bound to 10req/s. Thread pool won't be involved. And to increase throughput on a multicore system is to use other cores.
Then it comes asynchronous or callback functions. While it does give you a sense of parallelization, what really happens is it waits for the async code to finish in background so that event loop can work on another function call. Afterwards, the callback codes still has to run in event loop, therefore all your written code are still ran in the one and only one event loop, thus won't be able to harness multi-core systems' power.
The said document clearly states that Node is single-threaded:
A single instance of Node.js runs in a single thread. To take advantage of multi-core systems, the user will sometimes want to launch a cluster of Node.js processes to handle the load.
This way Node process has a single thread, unless new threads are created with respective APIs like child_process, cluster, native add-ons or several built-in modules that use libuv treadpool:
Asynchronous system APIs are used by Node.js whenever possible, but where they do not exist, libuv's threadpool is used to create asynchronous node APIs based on synchronous system APIs. Node.js APIs that use the threadpool are:
all fs APIs, other than the file watcher APIs and those that are explicitly synchronous
crypto.pbkdf2()
crypto.randomBytes(), unless it is used without a callback
crypto.randomFill()
dns.lookup()
all zlib APIs, other than those that are explicitly synchronous
A single thread uses 1 CPU core, in order to use available resources to the fullest extent and utilize multicore CPU, there should be several threads, the number of cores is used as a rule of thumb.
If cluster processes occupy 100% CPU and it's known there are other threads or external processes (database service) that would fight over CPU cores with cluster processes, the number of cluster processes can be decreased.
