On Fri, 2002-01-11 at 08:11, Jesus Cea Avion wrote: > > This timeout is not for message sent, but per "write" syscall. > > There is a problem, nevertheless: efficiency. If you do a "signal" and > two "alarm" per "write", your CPU usage will goes up. We should take > some experiments here to measure real impact. > > With a 512 "chunk" write buffer, for example, a megabyte email will > have, at least, 2048 "write" calls.
How about wrapping the entire transaction, not the individual writes? That would mean only one alarm and signal call for the entire megabyte.