Coroutine-based concurrency with gevent¶
Nucleon is tightly integrated with the gevent library, which makes it possible to perform multiple IO operations apparently concurrently, without the overhead of operating system threads.
This model of programming can require a different way of thinking. Specifically, unless a greenlet yields control either explicitly or by blocking on something, it will block everything else happening anywhere else in the nucleon application. For example an operation that takes 1 second to complete prevents any other requests from being served for a whole second, even if those operations would consume a tiny amount of CPU time. This could cause severe performance problems.
On the other hand, blocking network operations are very efficient. This includes database operations, REST calls, or communication with AMQP. gevent patches Python so that instead of blocking, other operations will be processed as if in the background. This includes Python operations that block, as well as pure Python libraries that use those operations.
Native libraries can completely block gevent. If an external library performs some blocking operation, your entire application will grind to a halt. You should identify whether the library supports non-blocking IO or can be integrated with an external IO loop, before attempting to integrate the library.
You might need to be particularly careful if a library performs I/O as a hidden side-effect of it normal operation. Some XML-processing libraries, for example, may make web requests for DTDs in order to correctly process an XML document.
You should also watch out for unexpected DNS activity.
To use gevent to best effect you should try to ensure that CPU is used in very short bursts so that the processing of other requests can be interleaved.
Diagnosing blocking calls¶
The Linux strace command can be used to print out the system calls used by a nucleon application.
$ strace -T nucleon start
The -T option will make strace display the time spent in each system call - pay attention to any calls with particularly large values, other than epoll_wait() (which is how gevent stops when all greenlets are blocked).
Undertaking more expensive processing¶
If you do need to use more CPU time very rarely, then it’s possible to mitigate the impact to other requests running at the same time.
The most direct way to do this is to explicitly yield control from within a greenlet. gevent will run any other greenlets that can run before returning control to the yielding greenlet. This is most similar to conventional threading.
A more elegant way to do this is to use a map-reduce model. In the map phase, a greenlet breaks up a task into many component tasks. These are each put onto a queue. Other greenlets pick up a task and execute them. The results are also put back into a queue. In the reduce phase some greenlet blocks waiting for responses and combines the results. Writing a task in this way can give extremely good scalability.
Nucleon runs in a single native thread in a single process, with all greenlets sharing the same memory space. Because of this, Nucleon apps can store data in application memory. No synchronisation primitives are required, so long as your application code never performs leaves the memory space in an inconsistent state while blocking IO operations are being performed.
Ensuring this is the case is preferable to using gevent.coros classes for locking, as this will simply reduce the number of greenlets eligible to run to completion while the greenlet holding the lock is blocked on I/O.