On Thu, 2002-02-28 at 22:30, Andrew Morton wrote: > There are a few things you can do: > > You'll want > http://www.zip.com.au/~akpm/linux/patches/2.4/2.4.18-rc1/make_request.patch > to fix the elevator starvation problem. This is officially in 2.4.19-pre1. > > You'll want > http://www.zip.com.au/~akpm/linux/patches/2.4/2.4.18-pre9/read-latency2.patch > which promotes reads ahead of writes. Dunno if this will ever be > merged, frankly. It's a *fundamental* change to what is probably > the single most important pieve of code in the kernel, from a performance > point of view. > > Now that's all well and good. With these patches, reads should > proceed nicely. But if the process which is trying to read things > accidentally does a write, it'll end up at the back of the > queue, delayed for a long time. You shuld prevent such accidents > by mounting your filesystems with the `noatime' option. I tried to apply the first patch to the RH kernel (first 2.4.9-21, which is the one I'm running, and then 2.4.17-0.18 from rawhide, which apparently is 2.4.18-rc1+rmap+O(1)+...), but they are significantly different in blkdev.h and ll_rw_blk.c and I'm a bit wary that the experience could turn bad (I would have to test on my work machine). So I think I'll wait till the thing advances a bit more. As for the 2nd patch I think it is a bit too radical for me... And the caveat about mounting with noatime is quite awkward btw. Just out of curiosity, wouldn't there a be way to avoid that kind of problem? Thanks a lot for your help, Diego -- ------------------------------------------------------- Diego Santa Cruz PhD. student Publications available at http://ltswww.epfl.ch/~dsanta Signal Processing Laboratory (LTS) Swiss Federal Institute of Technology (EPFL) EPFL - DE - LTS, CH-1015 Lausanne, Switzerland E-mail: Diego.SantaCruz@epfl.ch Phone: +41 - 21 - 693 26 57 Fax: +41 - 21 - 693 76 00 -------------------------------------------------------
