The problem was that with high load, circuit package window was
reaching 0. Whenever we got a circuit-level sendme, we were
reading a lot on each socket, but only writing out a bit. So we
would eventually reach eof. This would be noticed and acted on
even when there are still bytes sitting in the inbuf.
svn:r2932
directories on disk. Make authdirservers cache *their own*
directories instead of the ones they download. This gives better
liveness, and marginally better resistance to broken authdirservers.
Authdirservers don't cache running-routers at all: r-r isn't stored to
disk, so there's no point right now.
svn:r2893
they weren't writing anything into their datadirectory, so
when they stop and start, they would know nothing until they
fetched a directory from another authdirserver.
svn:r2891
- DirFetchPeriod for fetching full directory,
- StatusFetchPeriod for fetching running-routers,
- DirPostPeriod for posting server descriptor,
- RendPostPeriod for posting hidden service descriptors.
Also make sure the hidden service descriptors are at a random
offset from each other, to hinder linkability.
svn:r2889
Split logic to initiate dirfetch, running-routers fetch, and
descriptor post. arma: There are some XXXs here that raise design
questions which we should solve before the next release.
The biggest problem is this: Right now, the directory is about 50X as
large as running-routers uncompressed, and about 36X as large
compressed. Assuming:
- everybody gets the compressed version of everything,
- everybody gets cached directories from random dirservers and
uncached r-r from authdirservers
- everybody downloads r-r at the same rate they now download dirs,
then using r-r from will *increase* authdirserver directory bandwidth usage
if there are significantly more caches than authdirservers.
I think it's safe to leave this in for now, since there aren't 3x36 caching
dirservers, but we should make everybody with a dirport cache running-routers
soon. But I could be wrong.
svn:r2872
