Junliang, > Hello Honza, > > 在 2014-07-15二的 15:59 +0200，Jan Friesse写道： >> Junliang, >> >>> Hi all, >>> From this mail list , I found some guys had done many latency tests >>> with totem protocol. And in corosync configuration file , all parameters >>> are set in millisecond's(ms) level. I just wonder if it could be >> >> It's impossible directly from configuration file. And in a code such >> change would be quite difficult. If you want to try, take a look to >> qb_loop_timer_add calls (and specially QB_TIME_NS_IN_MSEC constant). >> > > Yes, it is quite difficult to change source code. What I really want to > do is finding out how to tune corosync in a small cluster (right now no > more than four nodes). > >>> possible for us to tune corosync in microsecond's(1/1000 ms) level, of >>> course, only for two nodes' case(connect two nodes directly without >>> switches or routers). Maybe it doesn't make sense to use corosync in >>> such "real time" case. >> >> Honestly, I don't think corosync will work with such a small timeouts. >> From time to time it's even problematic to create membership with >> default 1sec timeout. Also Linux is not very real time, and (for >> example) IO can kill "realtime" behavior. >> >> Just for curiosity, what are you trying to achieve? Having shorter then >> minimal 30 ms for failure detection or something else? >> > > Yeah, a shorter failure detection is my goal. Now we have a two-node > cluster, running same application with same inputs, and then produce > same output. Finally, we choose an output of one node to use and drop Ok > another node's output. It takes about 20ms to do HA switch (just tell > which node's output to use) while failures happen. Now, we try to > replace this mechanism with corosync and cluster resource manager like > pacemaker. Maybe it is a wrong way, but we want to have a try. By the > way, what would you do while in such a low switch time situation? > Basically, we didn't had such requirement before. And honestly, I'm unsure how to solve it (as I said, there are productions where even 1sec is not enough). At least corosync is really not designed for such use-case. But if you want, you can give a try to change MINIMUM_TIMEOUT (totemconfig.c file) to some lower value and you will see what will happen. Regards, Honza > Thanks, > Junliang Li > >> Regards, >> Honza >> >>> >>> Thanks, >>> Junliang Li >>> >>> _______________________________________________ >>> discuss mailing list >>> discuss@xxxxxxxxxxxx >>> http://lists.corosync.org/mailman/listinfo/discuss >>> >> > > _______________________________________________ discuss mailing list discuss@xxxxxxxxxxxx http://lists.corosync.org/mailman/listinfo/discuss
