On Sun, Mar 15, 2009 at 10:48:49AM -0500, Dan Nelson wrote: > In the last episode (Mar 15), cpghost said: > > I've noticed that when a huge, partially or totally swapped out process > > exits, there is a lot of disk activity going on, before the process truly > > dies. This is not necessarily due to sync(2), because it also happens > > with CPU bound processes that write very little output. > > > > Not sure what's really going on there, but apparently, the process reads > > in pages from swap that have been paged out previously (according to > > top(1)). > > Are you sure this is actually in _exit, and not in a cleanup function > executed by the application as it exits? If there is a large linked list, > for example, and the author has decided to actually free the list before > exiting instead of just letting it disappear when the process exits, each > swapped-out page will have to be brought back in as the list is traversed. > C++ programs may have destructors doing this behind the scenes.
Yes, that's quite possible. Meanwhile, I'm suspecting that free(3) is the culprit, and not the vm subsystem itself; though I was not yet able to construct a good example to be sure. And you're quite right: cleaning up paged-out linked lists or other dynamic data structures, either explicitly or via C++ destructors is also an obvious reason for swap activity. Didn't think of it first. > Best way to figure out what's going on is to attach to the program with gdb > while it's swapping, and print a stack trace. I'll try this. Very good idea. > Also, since you mentioned a "totally swapped out" process exiting, are you > terminating it externally with kill -9? It may be writing a core dump, > which will force the kernel to pull back swapped-out pages to write them to > the core file. Also an excellent point. I'm just killing them with Ctrl-C (SIGINT), which won't result in a core dump. But a core dump would also reawaken the pages, that's quite clear. > Dan Nelson > dnel...@allantgroup.com Thanks for all the hints. The fog is slowly lifting. ;-) -cpghost. -- Cordula's Web. http://www.cordula.ws/ _______________________________________________ firstname.lastname@example.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "freebsd-questions-unsubscr...@freebsd.org"