(Illustration by Gaich Muramatsu)
On Sun, Feb 11, 2007 at 03:32:59AM +0100, Davor Ocelic wrote: > On Sun, 11 Feb 2007 03:19:22 +0100 > Davor Ocelic <docelic_at_hcoop.net> wrote: > > > Systems will most probably be Intel 64bits. > > > > Even though I've got some specifications in my head, > > I have opened hands regarding the hardware to acquire > > for this purpose. I would appreciate all organizational > > hardware and software Coda-related suggestions. > > Forgot to mention; the whole system would go in production > in 5 or 6 months of which, things going well, there could be > two monts available for testing (primarily Coda). > > Are there some 64-bit issues or other stoppers that we could > get solved in that timeframe? (Possibly with some help from > this setup while it's in test phase.) 64-bit is definitely still an issue. At the moment a lot seems to work, mostly thanks to a large patch I got from Brett Lynmn which solved most of the really bad issues. On top of that I've managed to catch a lot of things that were flagged by the compiler. However there are still remaining problems which the compiler can't find. I think many problems are hidden by explicit casts that were added to solve compile issues in the Coda-4 to -5 timeframe. I actually got a client working on an Intel CPU with 64-bit extensions (EM64T) and committed the patches to a repository in Coda from that machine, so it seemed to work at first but the client crashed later on during a recursive ls. Another thing I noticed was that some character arrays that were written to the logs seemed to be off by 1 or 2 bytes (i.e. 2 random bytes followed by the expected string). I think there must be some strange alignment issue when a char [] is cast to a char * and back. I am not really an expert in alignment rules, but we probably have character arrays stored at non-aligned offsets in structures while a character pointer (string) is expected to be 32- or 64-bit aligned. In this case we probably end up with unexpected memory clobbers which is why there are seemingly random crashes. Since most of the remaining 64-bit problems are not always obvious, and I don't actually have a 64-bit machine sitting on my desk it will probably take a while before everything is resolved. For the rest your numbers seem quite reasonable, I think that our servers actually store more data. One possibility may be that your layout could hit a directory size limit. Coda directories cannot grow beyond 256KB, which I roughly estimate to about 4096 directory entries per directory. This is fairly low limit, the linux-kernel mailinglist stored as maildir can fill a directory in maybe 3 weeks or less. JanReceived on 2007-02-11 00:12:18