On Wed, 2018-12-12 at 23:11 +0100, Mkrtchyan, Tigran wrote: > Hi Frank, > > IMHO, very short leases can trigger a massive state recovery on > network hiccups. > > Our server offers 90 seconds to the client. The client usually renew > lease (sequence) once in a minute. > During high IO periods lease is not required and when idle, then once > in a minute is sufficient to > keep the mount alive. In a worst case, when a client got a lock and > dispersal, a competing lock/open > will block only for 90s. Whatever number you have, it should be > Given that TCP SYN timeouts generally follow an exponential law, and can easily reach > 20s, then I'd argue a 2s lease is on the extremely short end of tolerable. You probably want to do something a lot larger if you want lock recovery after a server crash to be reliable. > > BTW, IETF mailing list probably a better place for this question. > > Regards, > Tigran. > > > ----- Original Message ----- > > From: "Frank Filz" <ffilz@xxxxxxxxxx> > > To: "linux-nfs" <linux-nfs@xxxxxxxxxxxxxxx> > > Sent: Wednesday, December 12, 2018 10:14:59 PM > > Subject: What is a reasonable minimum lease time? > > We have an issue with the Ganesha server with very short (2 second) > > lease times. Ganesha uses a 1 second granularity for lease > > management, > > and considers a time since last renewed equal to the lease time as > > too > > long. The result is that the lease period may be short close to 2 > > seconds depending on when within a given second things actually > > happened > > (so a last renew at 0.99 with a subsequent renew at 2.01 which is > > just > > over one second looks like 2 seconds to Ganesha and thus is >= the > > 2 > > second lease time and not good enough. A simple change would be to > > change the >= to a >, which gives one more second, but it still > > could > > result in the lease time being almost 1 second too short which is > > significant with a 2 second lease time. But if the minimum > > reasonable > > lease time is more like 5 or 10 seconds, that 1 second becomes less > > significant. > > > > The bigger fix would be to use a finer grained time, but that adds > > complexity, but if people really want to run with 2 second lease > > times > > and it makes any kind of sense, we would need to make that change. > > > > Thanks > > > > > > Frank -- Trond Myklebust CTO, Hammerspace Inc 4300 El Camino Real, Suite 105 Los Altos, CA 94022 www.hammer.space
