You have multiple cores/procesors, use them
Async is best for doing heavy IO bound processing but what about heavy CPU bound processing?
The problem arises when single-threaded code blocks (ie gets stuck) on a long-running process. For instance, remember back when printing a word processor document would make the whole application freeze until the job was sent? Application freezing is a side-effect of a single-threaded application blocking during a CPU-intensive task.
In a multi-threaded application, CPU-intensive tasks (ex a print job) can be sent to a background worker thread thereby freeing up the UI thread.
Likewise, in a multi-process application the job can be sent via messaging (ex IPC, sockets, etc) to a subprocess designed specifically to process jobs.
In practice, async and multi-threaded/process code each have their benefits and drawbacks.
You can see the trend in the major cloud platforms, as they will offer instances specialized for CPU bound processing and instances specialized for IO bound processing.
Examples:
- Storage (ex Amazon S3, Google Cloud Drive) is CPU bound
- Web Servers are IO bound (Amazon EC2, Google App Engine)
- Databases are both, CPU bound for writes/indexing and IO bound for reads
To put it into perspective...
A webserver is a perfect example of a platform that is strongly IO bound. A multi-threaded webserver that assigns one thread per connection doesn't scale well because every thread incurs more overhead due to the increased amount of context switching and thread locking on shared resources. Whereas an async webserver would use a single address space.
Likewise, an application specialized for encoding video would work much better in a multi-threaded environment because the heavy processing involved would lock the main thread until the work was done. There are ways to mitigate this but it's much easier to have a single thread managing a queue, a second thread managing cleanup, and a pool of threads managing the heavy processing. Communication between threads happens only when tasks are assigned/completed so thread-locking overhead is kept to a bare minimum.
The best application often uses a combination of both. A webapp, for instance may use nginx (ie async single-threaded) as a load balancer to manage the torrent of incoming requests, a similar async webserver (ex Node.js) to handle http requests, and a set of multi-threaded servers handle uploading/streaming/encoding content, etc...
There have been a lot of religious wars over the years between multi-threaded, multi-process, and async models. As with the most things the best answer really should be, "it depends."
It follows a the same line of thinking that justifies using GPU and CPU architectures in parallel. Two specialized systems running in concert can have a much greater improvement than a single monolithic approach.
Neither are better because both have their uses. Use the best tool for the job.
Update:
I removed the reference to Apache and made a minor correction. Apache uses a multiprocess model which forks a process for every request increasing the amount of context switching at the kernel level. In addition, since the memory can't be shared across processes, each request incurs an additional memory cost.
Multi-threading gets around requiring additional memory because it relies on a shared memory between threads. Shared memory removes the additional memory overhead but still incurs the penalty of increased context switching. In addition -- to ensure that race conditions don't happen -- thread locks (that ensure exclusive access to only one thread at a time) are required for any resources that are shared across threads.
It's funny that you say, "programmers seems to love concurrency and multi-threaded programs in general." Multi-threaded programming is universally dreaded by anybody who has done any substantial amount of it in their time. Dead locks (a bug that happens when a resource is mistakenly locked by two different sources blocking both from ever finishing) and race conditions (where the program will mistakenly output the wrong result randomly due to incorrect sequencing) are some of the most difficult to track down and fix.
Update2:
Contrary to the blanket statement about IPC being faster than network (ie socket) communications. That's not always the case. Keep in mind that these are generalizations and implementation-specific details may have a huge impact on the result.
Concurrency and parallelism are two related but distinct concepts.
Concurrency means, essentially, that task A and task B both need to happen independently of each other, and A starts running, and then B starts before A is finished.
There are various different ways of accomplishing concurrency. One of them is parallelism--having multiple CPUs working on the different tasks at the same time. But that's not the only way. Another is by task switching, which works like this: Task A works up to a certain point, then the CPU working on it stops and switches over to task B, works on it for a while, and then switches back to task A. If the time slices are small enough, it may appear to the user that both things are being run in parallel, even though they're actually being processed in serial by a multitasking CPU.
Best Answer
The two most common problems in my experience of taking a class on the subject were the debugging of programs and the efficient distribution of resources.
Debugging a parallel program particularly on an independently managed thread system is astonishingly difficult. It's non-deterministic which means that your program may work 999 times out of a 1000 and that one time it fails just because something arrived in the wrong order or the thread manager didn't allocate correctly.
Atomicity errors are also common, this is why functional languages are becoming popular to handle concurrency as they wall off state (it's not that they don't have a state, it's that it's transparent).
The other big issue is communication. This falls into two categories: Resourcing and Time. Resourcing refers to often limited cache sizes and memory allocations for individual processes. Parallel systems often see use in data heavy applications, thus proper flow and packaging of that data is important. There's also the issue of dealing with the fact that there may not be a good algorithm that efficiently does what you want. Some tasks aren't not easily parallelizable and even if they are, the communication involved may render any speed gains moot. Time refers to the communication times between processors. This is less an issue in "multicore" systems than in distributed ones but it's still a problem.
You might ask why I'm talking about distributed systems at all. Well, some "shared memory" systems are actually distributed with particular facilities that make them function as shared memory.