Here's the big picture: ... periodically store some time-series numbers ... and with its own interval

As overviews go, that's a little like looking at the moon through the wrong end of a telescope. You kinds know what you're looking at, but the details decidedly fuzzy :)

A few things that would clarify:

• How is the shared hash keyed?

  Ie. Does seach thread write to a separate key (or separate set of keys)?

• What does "time-series" numbers" mean?

  An example would be good.

• When the sending thread wakes up to take it copy, does it just take whatever is available, or should it only take a "complete set"?

  What I'm getting at here is that threads are non-determanistic, so there is no guarentee that all your generating threads will have been given a timeslice by the scheduler between timeslices of your sending thread. If the time interval (controlled by a sleep?) is sufficiently long, the probability is that they will all get a timeslice in the interval, but it is unwise to rely upon such probabilities if the requirement is that the sending thread should always gather a complete set.

  So, I need to know if that is a requirement? And if it is, what steps are you currently taking to ensure that?

I'd need a far clearer picture of the task before I'd be prepared to offer a design for it's solution.