Thanks for the background. I'll look into it some more with different loads and tunings.
On 02/06/2013 03:41 PM, Richard Braun wrote: > On Wed, Feb 06, 2013 at 03:12:13PM +0100, Seth wrote: >>> On my workstation, and under the most extreme of loads I could think of, >>> the most I've seen reported 'Dirty" is 11MiB. >> (Note that it is usually solid 0M 90% of the time) > That's expected, and that's why I mentioned it's mostly useful for > system administrators and advanced users. The Linux kernel, and in > particular some recent versions, strongly limit the maximum dirty > memory [1]. Tuning vm.dirty_background_ratio and vm.dirty_ratio is the > most common way to optimize systems with high I/O load. > > For example, try setting them to respectively 25 and 75, do something > that generates many I/O requests, such as unpacking a large archive > (e.g. a linux tarball), and observe it on htop. You should also note > that the rest of your system should be smoother than with the defaults > (depending on your version; 2.6.38 already adds a lot in that regard, > but older kernels may simply block hard on I/Os until they're finished). > > You should now see how users who care about memory and I/O can find it > very useful. Compare that with the current separation of buffers and > cache. A lot of users just don't understand what the difference between > buffers and cache is, and can't make any useful decision based on that. > I personally don't get the point of statistics that don't help > understanding the work load and how to improve it. In addition, the > default life time of dirty pages in the kernel is 30 seconds, which > isn't that volatile. Think of a database server highly loaded with I/O > requests, the meter will just never be 0. > ------------------------------------------------------------------------------ Free Next-Gen Firewall Hardware Offer Buy your Sophos next-gen firewall before the end March 2013 and get the hardware for free! Learn more. http://p.sf.net/sfu/sophos-d2d-feb _______________________________________________ htop-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/htop-general
