On Thu, 22.07.10 10:42, Mike McGrath (mmcgrath@xxxxxxxxxx) wrote: > > how do you deal with circular dependencies in this case? > > I mean what will happen ? Will all services just deadlock? > > Malfunction ? Anything that guarantees correct initialization and > > behavior ? > > In addition to circular deps, have any studies been done on disk > contention when you just start everything all at once? If we're not > careful we could actually increase boot time in some scenarios. I guess > one way to check would be to go through all your init scripts and add a & > to the end of the start line? In theory the elevator should benefit from having more requests at the same time to choose from. However, theory is theory, and in fact the Linux IO scheduler does not benefit from this on rotating media too much right now. However, I think this kind of brokenness needs to be fixed in the kernel. This general problem has been discussed a couple of times in the meego community iirc. There have been suggestions to improve the situation. But nothing so far has been compelling and actually been fixed in the kernel. In generally I believe the right approach is to let the kernel IO scheduler schedule IO and not try to work around perceived brokeness simply by always serializing things on the userspace side already. Don't try to second-guess the IO scheduler, fix it. One could say that the IO scheduler is currently not directly optimized for this kind of IO load, because this load simply didn't exist before. It needs more optimization in this area. There have been attempts to work around this problem in userspace. For example with stuff like readahead. I personally believe that readahead is a deadend, and a correct fix for the problem would be more akin to Axboe's fcache, i.e. in order to optimize access time it is better to rearrange sectors then simply reading them in a different order and hoping this doesn't interfere with the normal IO scheduling applied. Looking at what Windows and MacOS do in this area is probably healthy. Both systems rearrange sectors on disk and parallelize as much as possible. I think that's bascially a good recipe we should follow too. systemd caters for the latter, we need kernel fixes to cater for the former. Lennart -- Lennart Poettering - Red Hat, Inc. -- devel mailing list devel@xxxxxxxxxxxxxxxxxxxxxxx https://admin.fedoraproject.org/mailman/listinfo/devel
