Messages in this thread | | | Date | Mon, 6 Apr 2009 15:17:28 -0700 (PDT) | From | Linus Torvalds <> | Subject | Re: [PATCH 0/8][RFC] IO latency/throughput fixes |
| |
On Mon, 6 Apr 2009, Ray Lee wrote: > > Your argument seems to be that someone may be relying upon default > kernel behavior and, at the same time, is willing to continually > upgrade their kernel.
No, I do think that the argument is valid per se: the crash behavior of "data=ordered" wrt "data=writeback" _is_ very different. And it's true that "data=ordered" has nicer behavior in the case of a crash.
But it is also true that with all the recent patches, "data=writeback" has become a _lot_ more attractive. It is slightly more ordered than it used to be, in a very particular way that makes one of the biggest downsides go away. And the latency improvements are really quite stunning (*).
So I think Hua's argument is real, but at the same time we should balance it against the fact that a lot of people will just use whatever is the default - and as a result we should strive to make the default be the thing that we think people would be happiest with.
I think "ordered" was a reasonable default, but that was at least partly because _both_ ordered and writeback sucked (partly in different ways).
I do think we could make it a config option.
Linus
(*) Admittedly with a very specific workload, but I suspect it's not that uncommon, and there will be people who never realized what their problem was and blamed jerky behavior on scheduler or something.
| |