Craig Shelley wrote: >With the noatime and nodiratime options, the problem of accessing files >causing a massive sync time has gone. See the results below. > >Although this option has made the system useable, I don't think it has >fully solved the problem. I think fsync() and sync() performance is >still poor, I will do a bit more testing and get back. > >Regards, > > > Sounds like we need to specifically optimize fsync. Well, we have known that for a while, but now it is more true.
Ways we can do it include using write twice journaling to a fixed journal area for fsync, and making it possible to specify for a particular file that it should have an atime, and then letting noatime be the default. Probably doing both is best. Atime has been hated by most fs designers for a very very long time, so there might be less resistance to a new standard for this than one would think. Hans