Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Wed, Dec 19, 2012 at 02:08:32PM +1100, Adam Goryachev wrote: > On 19/12/12 01:19, Prater, James K. wrote: > > Hello Andreas, > > > > Thanks for responding, below is a section of my current sysctl.conf file. As you can > > see that I used the alternative parameters that you had sent me. The resource spikes > > vary from 18-24 hours apart. The odd thing is that it sometimes occurs when the system > > is just idling i.e. 5-18Mbytes a second. Then it likely is caused by something completely different... > > # > > # Kernel tweaks to close the gap between the buffercache and DRBD DIO to remote store. > > # > > vm.swappiness = 10 > > vm.zone_reclaim_mode = 1 > > vm.dirty_background_ratio = 1 > > vm.dirty_ratio = 1 > > vm.dirty_expire_centisecs = 300 > > vm.dirty_writeback_centisecs = 60 > > vm.vfs_cache_pressure = 40 > > > > The settings forces the delta between data coming into the file server and data being set to the peer to be > > 1-3 seconds as opposed to nearly 25seconds. > > > > Shouldn't these two values be different: > > vm.dirty_background_ratio = 1 > vm.dirty_ratio = 1 > > ie, vm.dirty_ratio should be higher than vm.dirty_background_ratio. Not necessarily. The other way around (ratio smaller than background_ratio) can be beneficial as well, e.g. if you expect mostly streaming writers. Though I admit that using the same value for both is not very useful. Depends very much on the expected workload. (Same goes for the _centisecs: tuning up rather than down may help as well... it depends...) Also, please use *_bytes (not _ratio) ((unless you have an old kernel that does not yet know about *_bytes)), because even 1 percent of 120 G is a bit much for many storage backends. -- : Lars Ellenberg : LINBIT | Your Way to High Availability : DRBD/HA support and consulting http://www.linbit.com