Incidentally, the basis of my code was straight from the documentation for threads and threads::shared. Who knew I would be so off base?

Hm. Unless there is some new or modified documentation kicking around that I've not seen, I think you have misinterpreted the documentation. I am not aware of any docs that suggest the queue architecture you are using.

Namely: each worker has its own work item queue and a work item request queue. And in order to get a work item, it has to enqueue a request (its thread id) onto the request queue; wait (many time-expensive context switches) for the thread at the other end of that queue; to dequeue that tid and then enqueue a work item; and then continue waiting until the initiating worker gets another time-slices (many more context switches) before it can actually do some work.

The problems with this -- as I tried to analogise above -- are: a) you have many worker threads all waiting on one non-worker queue; b) one non-worker queue having to satisfy the demands of many worker threads.

Are you saying that the "guard" should start polling a single queue and stuffing things into it on demand?

No. One of the purposes of queues is to avoid polling. (IMO there should not be a dequeue_nb() method for a queue; but that's a different discussion :)

The non-worker threads job is to keep enough work in the (single) work-item queue to ensure that whenever a worker needs a work item, there is one immediately available. It doesn't have to ask for one, it just grabs one.

Its only other responsibility is to ensure that it doesn't overfill the queue. The 'damage' this does is to use more memory than is necessary. Basically whenever it (the non-worker, producer thread) has a new work item available, it checks the size of the work-item queue, and if it is bigger than some threshold value (2 or 3 times the number of workers is a good starting point), it simply sleeps for a short period to allow the worker threads to diminish the queue a little further.

Although this may sound like polling, it is quite different in that there is no single event or synchronisation involved. That is, it is not waiting for the queue length to fall to some exact number; but rather just until it falls below some number. Provided that number is sufficient to ensure that no worker is ever blocked waiting for a new item, this ensures that the workers run at their full potential speed without risk memory overruns by over stuffing the queue.

What to do... Could you steer me in the way of an implementation like the one you suggest?

The first thing is that all your workers should be reading from a single queue. After all they are all identical, so it doesn't matter which worker processes which work item.

That does away with the separate, per-worker item queues and the requests (tid) queue. The non-worker just stuffs every item it gets onto the common worker queue as it gets them; and only pauses briefly -- a single timeslice period ~10ms is a good starting point -- iff the queue is starting to overfill. If you get the length of the queue right, this will be a rare event.

This way each thread can -- assuming the nature of the work allows it -- process as many work items as it can in each timeslice it gets. Thus you maximise the timeslice utilisation and minimise the expensive context switches. This architecture also ensures that when some workers run more slowly -- processing bigger files -- the other workers can continue to run flat out if they get a bunch of small ones. Ie. The single queue/multiple workers scheme becomes self-load-balancing.

Your 7 step overview sounds reasonable except for the multiple queues and work item counting. Fix that and it should improve your throughput somewhat.

But overall, I think this is not a particularly good task for multi-tasking -- whether by multi-threading, multiprocessing, user-space threads (coroutines) or any other mechanism.

Ultimately, the whole thing is IO-constrained -- including the digesting of the files; you can only digest the blocks as fast as you can read them from disk and the time taken to read a (chunk of a) file from disk will always be many times greater than the time required to calculate the digest.

And if your multi-threading scheme means that you end up having different threads reading different files from the same physical spindle concurrently, you will actually slow things down because the disk head will be seeking back and forth between different files (thus disk tracks) trying to keep all your threads supplied with data.

And if you try to avoid that by using a semaphore to allow only one thread at a time to read from disk, you've effectively serialised the entire process and will see little or no gain from the threading.

I'm afraid there is simply no simple way to achieve substantial throughput gains from multi-tasking this type of IO-bound processing unless the filesystem being processes is itself distributed. That is to say: unless the filesystem is located on a big box of simple disks (eg. a RAIDed NAS, SAN or similar), then you'll find it very hard to achieve much better throughput than a well written, single-tasking, single pass through the filesystem gathering the stat info; and then a second, single-tasking, single pass digesting just those files that are still potential duplicates after eliminating the obviously different ones.