Thank you all for your replies and discussion!
I went ahead and implemented the one-thread data management queue. Much to my surprise, however, it didn't result in any perceived change in memory usage! It looks like Perl does some intelligent things in copying shared structures, probably something like Copy-On-Read.
Because of this, I then expected my long-term memory usage to go down. But after 10 hours of run-time that didn't seem to happen either...
I assume my usage pattern of the structure made it resistant to memory bloat. The structure is a kind of work queue that threads take work items from. They go through the structure periodically to check for valid elements, but this accesses only a small part of the data. When the rest of the data is read, it is discarded shortly after.
Now it seems the worst cause of bloat are the modules copied for each thread. I have minimized the number of use'd modules, but they still amount to about 5 MB per thread. With 200 threads this uses half the available 32-bit address space...
Optimizing modules goes out of the scope of my question, so I'm not going to dwell on that further. Suffice it to say that I have options there as I am using LWP::UserAgent which can be replaced by other modules like LWP::Parallel or HTTP::Async or even my own socket code.
In reply to Re: Sharing large data structures between threads
by Anonymous Monk
in thread Sharing large data structures between threads
by Anonymous Monk
| For: | Use: | ||
| & | & | ||
| < | < | ||
| > | > | ||
| [ | [ | ||
| ] | ] |