Messages in this thread | | | Date | Sat, 26 Jun 1999 18:40:59 -0400 (EDT) | From | Alexander Viro <> | Subject | Re: (reiserfs) Re: File systems are semantically impoverished compared to database |
| |
On Sun, 27 Jun 1999, Hans Reiser wrote:
> > Each layer with its own pointer structure adds additional pointers > because in a single tree system the depth of accesses is log N, but in a > two tree system the depth is 2 log(N/2), which is much more than log N.
WHAT??? 2 log(sqrt(n)), please (assuming as you apparently did that depth is divided evenly).
Hans, sorry, but it's plainly bogus - "trees" in structured files/bundles/whatever are attached to *leaves* of the large tree. And if you are thinking of the allocation metadata of the underlying file - how deep is it going to be? You are spending most of the time doing actual file data IO, not the tree traversals. And data IO is *not* going to see the intermediate level at all.
> Each layer of interface adds a layer of functionality distortion, and > the capabilities of a two layer system is in some ways the intersection > not the union of the separate capabilities of the layers.
Translate it, please.
> Recovery is necessarily much uglified, more than twice as ugly, and this > makes performance worse since performance normally has recovery as the > primary crippling factor that prevents the use of various desired > algorithms. > > Hans > > Alexander Viro writes: > > > > > > On Sat, 26 Jun 1999, Hans Reiser wrote: > > > > > > > > The goal of reiserfs is not to get every person, who has invented yet > > > another namespace that can share the interactions it conducts with no > > > other namespace, to convert to reiserfs. The goal of reiserfs is to one > > > by one eliminate the reasons why these new namespaces keep getting > > > invented rather than using the filesystem namespace. I cannot > > > unify the namespaces, I can only somewhat reduce the reasons why they > > > fragment, and pontificate a bit at those who fragment them. > > > > > > Two filesystem layers are indeed terrible. What you are doing is not > > > analogous to, say, putting a file system ontop of a disk driver which is > > > ontop of a piece of hardware, no, you are writing a C interpreter in > > > Pascal. You are putting a motorcycle seat ontop of a truck chassis, and > > > then you are asking me to define why a truck chassis isn't as functional > > > as a motorcycle chassis, and given that you have already built a truck > > > chassis, you ask why it is that I think you won't have all the > > > functionality of a motorcycle using your motorcycle seat and handlebars > > > ontop of a truck chassis. What you want to do is like telling the TCP/IP > > > guys when they wrote TCP/IP that they should be using the serial line > > > driver as the bottom layer of their TCP/IP implementation, putting the > > > TCP/IP stuff in user space, and then preaching about the benefits of > > > increased portability to other Unixes that would result from that. > > > > Very colourful. What about a couple of technical arguments instead of > > Datamation-level analogies? >
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |